RoboPapers

Delta Intelligence’s Δ₀ is a humanoid full-body foundation model, which can perform long-horizon, whole body loco-manipulation. They show a single, generalist policy which can control a 69-DoF humanoid to perform everyday tasks at nearly human speed — using whole-body human data for large-scale pretraining.Xiaojian Ma and Siyuan Huang tell us more about how they achieved this, and about the importance of jointly designing policy and controller. Delta uses a delta action space, for example, which makes it easy to provide human interventions without interrupting the whole body controller: a human operator can correct the robot seemlessly while it works.They show a wide range of impressive humanoid loco manipulation tasks using their generalist model.Learn more on Episode 115 of RoboPapers, with Michael Cho and Chris Paxton!Learn MoreBlog post: https://deltai.com/en/blog/delta-0TranscriptThis transcript has been edited for clarity and readability. Repetitions, false starts, and incidental screen-sharing chatter have been removed.Introducing Delta IntelligenceMichael Cho: Welcome to another episode of RoboPapers. Today we’re glad to have Xiaojian and Siyuan. Xiaojian is the CEO and Siyuan is the chief scientist of a new startup called Delta Intelligence. They’ve just released their model, Delta-0, along with a detailed technical blog post.We’re very glad to have both of you here to share your work. To start, could you each tell us a little about yourselves? Xiaojian, perhaps you can go first.Xiaojian Ma: Thanks, Michael and Chris, for the invitation. I’m Xiaojian, the CEO and founder of Delta Intelligence. Delta is dedicated to solving general intelligence for humanoid robots. We’re focusing on bipedal robots because we believe humanoids are the future and a general form of physical intelligence.We’re building foundation models that include both the brain—the large models—and the controller. We also work on hybrid force-position control for humanoids. Delta-0 is our first foundation model, showcasing the capabilities we currently have on Unitree G1 robots. We’re expanding beyond Unitree to other robot vendors in China and elsewhere, because we want our model to have cross-embodiment capabilities.Before founding Delta, I was a researcher in machine learning and robotics. I returned to China to join BIGAI as a research scientist in embodied AI. Before that, I was a PhD student at UCLA. During my PhD, I did research internships with teams at Google Robotics, NVIDIA Research, and Google DeepMind.My projects mostly involved building large multimodal models for reasoning and embodied decision-making. That’s the background I’m coming from.Siyuan Huang: Thank you for the invitation, Chris and Michael—two old friends. We’ve shared papers with you and my students several times, but this is the first time we’re discussing a blog post about Delta-0, our new foundation model.I’m the chief scientist of Delta. We released this blog at the end of last month, and it has prompted discussion about how to approach humanoid intelligence and build foundation models for whole-body loco-manipulation.We’re excited about the development of humanoid foundation models. This blog summarizes how we think we can build better models and address the coupled problems of the controller, hardware, and model.Watching the DemoSiyuan Huang: Let’s go over the demo first. We have two robots in an indoor environment. The first is doing what we call whole-body dexterous manipulation. The hardware is a Unitree G1 with two WUJI hands. The first task is putting a record onto a turntable.Michael Cho: And doing a dance, too. Nice.Siyuan Huang: Our expectation is that, in the future, we can have robots in our homes doing the same kinds of things people do: putting on a record, listening to music, doing a little dance, and taking a rest.They can also do everyday housework, such as making the bed, picking up a game controller, and arranging the bedding. These are long-horizon loco-manipulation tasks. They require whole-body coordination to finish everyday activities, such as picking up a toy from the ground and putting it on the bed.The robot can also interact with articulated objects, applying force to open and close doors. Here it’s opening the dishwasher and using its left hand for support so it can apply more force.Chris Paxton: That’s really cool. It was bracing itself against the fridge and pulling with the other hand.Xiaojian Ma: That’s right. It’s like using your core to adjust your body and pull with more force.Chris Paxton: It’s very human, very fluent.Siyuan Huang: Here it checks the food in the refrigerator—and takes a rest as well.Michael Cho: This is so cool. It’s really long horizon.Chris Paxton: It’s a fantastic video. I also want to note how few cuts there were. Was there just one, when you switched from the first robot to the second?Siyuan Huang: There are actually two cuts.Michael Cho: The second part is continuous. All this opening of doors is nontrivial. Very impressive. Please take us through it.Siyuan Huang: The video has two parts. The first shows very dexterous manipulation. The second shows very long-horizon loco-manipulation: about three minutes of uncut video, with nine atomic tasks and more than thirty steps. It’s challenging for the robot to finish all of those.We’ve seen companies such as Figure show videos like this. As Michael said, opening and closing objects requires a lot of interaction with articulated objects. Connecting those tasks is particularly hard because each task starts from wherever the previous one ended.We don’t use any SLAM-based navigation skill in this video. The model drives navigation, upper-body interaction, and lower-body interaction. All of that comes from one model.Why Loco-Manipulation Is HardSiyuan Huang: First, we want to explain how we think about humanoid foundation models and why humanoid intelligence matters.The humanoid form offers the possibility of working wherever people do. We think it’s a general form that can work in more places than a wheeled robot. Humanoid intelligence turns that possibility into useful, reliable actions.Right now, we’ve seen a lot of demonstrations. In the future, we want to turn those capabilities into deployments. That’s the vision of Delta Intelligence.One challenge is long-horizon loco-manipulation. We expect the robot to perform extended activities, rather than just individual pick-up actions in front of a table. Every action changes the conditions for what comes next: the surroundings change, the robot’s location changes, and the mode of interaction changes. It’s difficult to ensure stability throughout a long-horizon task.We see loco-manipulation as one of the most challenging unsolved problems in robotics, and a core problem for humanoid intelligence. There are several reasons.First, the robot must control a moving body while perceiving and manipulating a changing world. That’s different from manipulation alone. Previously, we often tried to separate locomotion and manipulation.For locomotion, an RL-based controller can make the robot dance or perform extreme movements. We’ve discussed impressive examples on RoboPapers, such as OmniXtreme, where a single policy produces a broad range of movements. That’s locomotion and whole-body control.For manipulation, you can have a fixed base and use a gripper or dexterous hand. But fixing the base limits the robot’s action space. We want a larger action space, more like a human’s, so we combine locomotion and manipulation. The robot needs to move while perceiving and interacting with a changing world. That creates a large action space and a difficult coupled problem.Second, each action determines the starting conditions for the next. That’s one reason long-horizon tasks remain a holy grail of robotics. A large language model can decompose a task into atomic subtasks, but physically connecting them requires precise positioning and adjustment. The robot needs to tr… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Robotics has a data problem, but not the one you’re thinking of: almost all robotics data is of successful instances of robot behaviors. But many of the most exciting research areas for robotics today, including world models and reward model learning, require failures: interesting examples of manipulation trajectories that did not succeed.Claas Voelcker, Arpit Bahety, and Paul Zhou join us to talk about their data collection effort and their upcoming workshop on “oopsie data”: all these valuable failures that are currently being thrown out. Their goal is to collect the first large community-driven open-source dataset of failure examples.Watch Episode 114 of RoboPapers, with Chris Paxton and Jiafei Duan, to learn more!Learn MoreProject site: https://oopsie-data.com/ This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Using simulation to teach robots would unblock the massive scale of robotics data that seems to be missing right now. The issue is usually sim-to-real transfer; MolmoBOT aims to solve this with sufficient diversity of simulation data, so that no real data is necessary for robot deployment. It has a few parts: MolmoBot-Engine produces diverse pick-and-place demonstrations across many different simulated environments; MolmoBot-Data is a dataset of 1.7 million expert trajectories for item manipulation and pick-and-place tasks; and they train a number of VLA policies which can achieve strong performance on real-world pick-and-place tasks with zero real robot data. Rose Hendrix, Snehal Jauhri, Abhay Deshpande, and Maya Guru join us to give us the details.To learn more, watch RoboPapers episode 113 with Chris Paxton and Ruijie He!AbstractA prevailing view in robot learning is that simulation alone is not enough; effective sim-to-real transfer is widely believed to require at least some real-world data collection or task-specific fine-tuning to bridge the gap between simulated and physical environments. We challenge that assumption. With sufficiently large-scale and diverse simulated synthetic training data, we show that zero-shot transfer to the real world is not only possible, but effective for both static and mobile manipulation.We introduce MolmoBot-Engine, a fully open-source pipeline for procedural data generation across robots, tasks, and diverse simulated environments in MolmoSpaces. With it, we release MolmoBot-Data, a dataset of 1.7 million expert trajectories for articulated object manipulation and pick-and-place tasks. We train three policy classes: MolmoBot (a Molmo2-based VLM with a flow-matching action head), MolmoBot-Pi0 (replicating the π₀ architecture for controlled comparison), and MolmoBot-SPOC (a lightweight policy for edge deployment). Without any real-world fine-tuning, our policies achieve zero-shot transfer to unseen objects and environments, reaching 79.2% success rate on real-world tabletop pick-and-place, outperforming π₀.₅ at 39.2%.Learn MoreProject Page: https://allenai.github.io/MolmoBot/ArXiV PDF: https://arxiv.org/pdf/2603.16861TranscriptProduced by ChatGPT and Riverside.Introducing MolmoBotChris Paxton: Hey everybody, and welcome to another episode of RoboPapers. Today we’re excited to have the team behind the MolmoBot paper here to talk about their work. Would you like to introduce yourselves?Rose Hendrix: I’m Rose Hendrix. Until recently, I was a research scientist at AI2, and now I’m at a startup.Abhay Deshpande: I’m Abhay Deshpande. I conducted this work while I was a predoctoral researcher at AI2, and now I’m a PhD student at Princeton University.Maya Guru: I’m Maya Guru. I’m currently a predoctoral researcher at AI2, and I’m looking for PhD positions for this fall.Chris Paxton: Awesome. Would you like to get started and tell us a little about what you’ve been working on? Rose, maybe you can share your screen?Abhay Deshpande: I’ll pull it up.Rose Hendrix: Thanks, Abhay. On a previous episode of RoboPapers, the MolmoSpaces team came on to talk about their work, which is closely related to this. MolmoSpaces is about scaling synthetic environments and evaluation. MolmoBot is about cashing out that potential: scaling synthetic data generation, then training policies on that data to achieve zero-shot sim-to-real transfer for manipulation.MolmoBot is a data engine and a family of policies. We show that the policies transfer in large-scale simulation evaluations and in the real world, both for tabletop pick-and-place and for mobile manipulation on the RB-Y1 platform. That includes opening doors, drawers, and cabinets, as well as object manipulation and pick-and-place.Generating synthetic robot data at scaleAbhay Deshpande: As Rose mentioned, we have this large framework, MolmoSpaces, for synthetic data and large-scale simulation. Within that, we built MolmoBot-Engine, a data engine for generating extremely diverse, large-scale datasets.The basic idea is to roll out heuristic demonstrators in these environments, then apply very aggressive randomization: camera extrinsics and intrinsics, textures, lighting, colors, and all sorts of other things. You keep rolling that out and scaling with compute, and you get high-throughput, large-scale data.For a concrete example, you sample an arbitrary scene from MolmoSpaces. There are more than 100,000 scenes available. Then you sample a workspace, choose an object for a pick-and-place operation, place a receptacle for the object, and put the robot in the workspace.From there, we randomize lighting, the robot’s initial configuration, inertial and frictional parameters, and other aspects of the scene. Then we roll out a scripted demonstrator. Since we have access to privileged information in simulation, a fairly simple heuristic policy can perform well under a wide variety of conditions.Of course, there are so many things we could randomize: textures, lighting, robot configurations, camera poses, task instructions, templates, and much more. How do we decide which ones matter? What we found is that a good philosophy is basically: yes. If you can randomize it, you generally should. That extreme diversity goes a long way toward making policies robust in the real world.We generate several different modalities. By default, we have two external cameras and a wrist camera. Since we’re using the DROID setup, we replicate those cameras in simulation: ZED 2 cameras for the external views and a ZED Mini for the wrist camera. We also have a randomized GoPro view.The GoPro has a very wide field of view and significant fisheye distortion. We apply post-processing to warp the rectified image so it looks like what the real GoPro would produce.In addition to RGB and depth, we save the camera intrinsics, since we’re randomizing them. We also save task-relevant object keypoints, which lets us compute visibility heuristics or do image-space reasoning downstream.For actions, we save all of the action representations rather than prescribing just one. That way, people using MolmoBot-Data can support a wide variety of downstream applications. We also save proprioception, joint positions and velocities, end-effector poses and twists, and other task-relevant privileged state: the demonstrator’s current phase, contact forces and points, grasps, door states, and so on.Finally, we label everything with rewards, success indicators, scene metadata, task prompts, and language referring expressions.Chris Paxton: One question: all these grasps and door states you’re talking about—where do those come from? Are those heuristics?Abhay Deshpande: Generally, they’re derived from the scripted heuristics or read directly from simulation. A door state might just be a joint angle, which we can access directly. The same goes for contact points and forces. Others, like rewards and success, are based on predicates: conditions that need to evaluate to true for us to say the task succeeded.Rose Hendrix: This was also motivated by the fact that knowing what the demonstrator is trying to do can be a useful training signal. That’s a strength of synthetic data. If you’re collecting human teleoperation demonstrations, it isn’t necessarily possible to extract that kind of nuanced information about what the operator intended at any given moment.RJ He: Was all the data scripted, or were there other sources?Abhay Deshpande: It was all scripted. We found that fairly simple heuristic policies go a long way. We have around 1.7 million demonstrations, and they’re all generated with scripted policies. By randomizing enough dimensions, you get surprisingly far.You can see in the data distribution that the scripted motions sometimes look a bit robotic—which, of course, is fitting for a robot. But we didn’t encounter major downstream problems, and the policies performed well in the real world. It might become more of an issue when co-training with dat… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Touch is crucial for humans when they want to perform precise tasks to a high degree of reliability, and yet robots almost universally use vision — they’re often completely blind to this whole modality. Two works from Tong Jian, Tao Yu, Saleh Nabi, and Matteo Grimaldi show us how important this sense is. In SlipSense, they look at how to detect objects slipping in grippers using a compact multimodal tactile sensor, which generalizes across unseen objects and to different robots. They then talk about how tactile sensing can be used to enable high real-world performance with a challenging cable-tracing task, where a cable must be fed through the hand with a repeated pinch-and-curl motion.To learn more, watch Episode 112 of RoboPapers today, with Michael Cho and Ruijie He!SlipSense AbstractSlip detection is fundamental to dexterous manipulation, yet existing systems often lack precise characterization of detection latency and cross-platform generalization. We present SlipSense, a multimodal tactile slip-detection framework built on TacV5, a compact sensor integrating a 32×32 piezoresistive array operating at 240 Hz and a 3-axis MEMS accelerometer operating at 8 kHz. The piezoresistive array captures spatial pressure distributions, while the accelerometer captures friction-induced vibrations, providing complementary slip cues. The framework performs modality-specific encoding, intra-sensor fusion, and cross-modal attention with causal temporal prediction at 240 Hz. Experiments on a dataset of 1.4 million frames spanning 37 objects demonstrate the complementarity of the two modalities. SlipSense achieves 96.7% Macro F1 with a false-positive rate below 1.6%, detecting 76% of slip events within 23.1 ms. When trained solely on UMI data, SlipSense generalizes zero-shot to a Tesollo dexterous hand, transferring across unseen objects, distinct sensor units, and robotic platforms without retraining.Touch2Trace AbstractDexterous manipulation of deformable objects demands continuous fingertip-level regulation of pressure, friction, and incipient slip. We study one of the most challenging cases: dexterous cable tracing, feeding a cable through the hand with repeated pinch-and-curl motions of the thumb and index finger. We introduce Touch2Trace, a tactile-driven imitation-learning system for this task, and provide, to our knowledge, the first systematic real-world characterization of how encoder pretraining, control rate, temporal context, and spatial resolution each shape policy performance. The winning learning recipe combines a tactile encoder pretrained for a custom 32 x 32 piezoresistive sensor (TacV5) via self-supervised learning with a lightweight transformer policy trained on teleoperated demonstrations via behavior cloning, deployed at 60 Hz on a Tesollo DG-5F hand. Tactile feedback without vision or explicit cable-state estimation significantly improves tracing performance versus a proprioception-only baseline: from 0.2 cm to 20.1 cm mean distance and 0% to 93% success rate, with zero-shot transfer to unseen cables and routing conditions. The results quantify the influence of key parameters in tactile-driven systems for reliable dexterous deformable object manipulation.Learn MoreSlipSense Project Page: https://tjian-ai.github.io/corl26-slipsense/SlipSense on ArXiV: https://arxiv.org/abs/2609.15910Touch2Trace on Project Page: https://grmmtt.github.io/corl26-touch2trace/Touch2Trace on ArXiV: https://arxiv.org/abs/2609.15921TranscriptIntroductionMichael Cho: Hello, everyone. Welcome to another episode of RoboPapers. Today we’re really happy to have the crew from Analog Devices, the Dexterous AI Group: Tao Yu, Tong Jian, Saleh, and Matteo. They’re coauthors of two papers that they’ll walk us through. One is Touch2Trace: Tactile-Driven Imitation Learning for Dexterous Cable Tracing, and the other is SlipSense: Multimodal Tactile Learning for Low-Latency and Generalized Slip Detection.Thank you so much for coming, guys. Maybe we can start with everyone introducing themselves, and then go through the papers. Tao, do you want to start?Tao Yu: Yeah, great. Thanks. My name is Tao Yu. I lead the DexAI team, short for Dexterous AI Group, at Analog Devices. The team was set up about two years ago, and our mission is to unlock humanoid dexterity. That’s how we started looking into tactile sensing, which we believe is one of the missing pieces in dexterous manipulation.With me are the other team members. Maybe Tom can go next.Tong Jian: Yeah, I’m Tom. I’ve been with the DexAI team for two years, working on applied machine learning and, more recently, robot learning.Saleh: Matteo, do you want to go next?Matteo Grimaldi: Hello, everyone. I’m Matteo. I’ve been working with the DexAI team for two years now, mostly on robot learning.Saleh: And I’m Saleh. I also joined Dexterous AI about two years ago. When I met Tao, we were still prototyping a lot of things on the sensor side. The two works you’ll see today are the result of a lot of interesting internal research, both on the hardware side and on the algorithm side.Tao Yu: I also want to add that I personally believe this isn’t just a hardware problem or an algorithm problem. It’s a combination of both. One question we’ll touch on in these papers is what hardware capabilities can unlock learning, and vice versa: what can learning tell us about the hardware requirements?We believe the path forward is a loop: create new hardware, integrate it with a robot, train a model, demonstrate whether the new hardware capability improves performance, and then use what we’ve learned to design the next generation of hardware.Michael Cho: Very cool. I’m really looking forward to these two papers. Which one do you want to start with?Tong Jian: Maybe we can start with SlipSense. Let me share my screen.Michael Cho: Okay, cool.SlipSense: Detecting Slip Quickly and ReliablyTong Jian: Can you see the screen? Great. We’ll be talking about two works accepted at CoRL 2026, both built on our TacV5 tactile sensor. They demonstrate tasks where vision alone isn’t enough.I’ll begin with SlipSense, and Matteo will follow with Touch2Trace. As the title suggests, SlipSense studies slip detection. When an object starts to slip from a robot gripper or fingers, we want a model that can detect it quickly and reliably.Slip detection has a long history. Researchers have been working on it for decades, and the core problem is well understood. But two things remain relatively underexplored. The first is detection latency. A lot of prior work reports accuracy without really characterizing latency.To emphasize, detection latency is not the same as inference speed. It’s the time between when slip actually begins—the slip onset—and when the model first detects it. That’s the interval between these two lines on the plot. A model might run in two milliseconds, but if it needs 100 milliseconds of evidence to make a correct prediction, that’s still 100 milliseconds of detection latency. By the time the model detects slip, the object may already be gone.That’s why we care so much about detection latency. It depends both on the sensor’s capabilities and on the model’s capabilities.RJ He: Can I jump in here? This might be a naive question, but it sounds like you’re saying that different materials and different objects can actually produce different slip-detection latencies. Is that correct?Tong Jian: Yes. It depends on many things: the material, the object, the actual slip speed, how quickly the sensor can capture the relevant slip features, and how quickly the model can interpret those features and make the correct detection. All of those things matter. There’s not just one dominant factor.RJ He: I see. And my assumption is that this is a learned detector, not just a property of the sensor?Tong Jian: Yes, exactly. It’s learned. And that brings us to another important aspect: detection latency naturally trades off against false alarms. It’s a sensitivity-versus-specificity… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Reinforcement learning is often seen as key to achieving scalable, highly reliable real-world robot performance, yet it requires good reward functions that can supervise the wide range of tasks that need to be performed. The obvious thing to try would be a large vision-language model (VLM), but these models often fail under common real-world conditions like partial observability and occlusion.As a solution Philip Schroeder and Thomas Weng propose SOLE-R1, a video-language reasoning model trained to serve as a reward signal for online RL. Given real video, SOLE-R1 will reason over input video frames, generates reasoning traces, and predicts a reward signal. This allows online RL from random initialization on new manipulation tasks.To learn more, watch Episode 111 of RoboPapers, with Michael Cho and Jiafei Duan!AbstractVision-language models (VLMs) have shown impressive capabilities across diverse tasks, motivating efforts to leverage these models to supervise robot learning. However, when used as evaluators in reinforcement learning (RL), today's strongest models often fail under partial observability and distribution shift, enabling policies to exploit perceptual errors rather than solve the task. We introduce SOLE-R1 (Self-Observing LEarner), a video-language reasoning model explicitly designed to serve as the sole reward signal for online RL. Given only raw video observations and a natural-language goal, SOLE-R1 performs per-timestep spatiotemporal chain-of-thought (CoT) reasoning and produces dense estimates of task progress that can be used directly as rewards. To train SOLE-R1, we develop a large-scale video trajectory and reasoning synthesis pipeline that generates temporally grounded CoT traces aligned with continuous progress supervision. This data is combined with foundational spatial and multi-frame temporal reasoning, and used to train the model with a hybrid framework that couples supervised fine-tuning with RL from verifiable rewards. Across four different simulation environments and a real-robot setting, SOLE-R1 enables zero-shot online RL from random initialization: robots learn previously unseen manipulation tasks without ground-truth rewards, success indicators, demonstrations, or task-specific tuning. SOLE-R1 succeeds on 24 unseen tasks and substantially outperforms strong vision-language rewarders, including Robometer, RoboReward, ReWiND, GPT-5, and Gemini-3-Pro, while exhibiting markedly greater robustness to reward hacking. We release all models, data, code, and demos at the anonymous page: this https URLLearn MoreProject page: https://philip-mit.github.io/sole-r1/ArXiV: https://arxiv.org/abs/2603.28730Transcript Transcripts are automatically generated by Riverside and are lightly edited by ChatGPT.Michael Cho [00:06]Hello guys, welcome to another episode of RoboPapers. Today we are really happy to have Philip Schroeder and Thomas Weng, two of the co-authors of this paper, SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning. Thank you so much for making the time, guys. Maybe as a start, Philip, you can share a bit about yourself and then Thomas.Philip Schroeder [00:27]Yeah, so my name is Philip. I’m a final-year PhD student at MIT in computer science. My work has mainly focused on improving the reasoning capabilities of language and vision-language models in challenging embodied settings. And over the last couple years, my work has mainly focused on reward reasoning over video, which is what we do for this paper.Thomas Weng [00:55]Cool. I’m Thomas. I’m a research scientist at RAInstitute. And more recently my work has been focusing on the pieces required to achieve like post-training and deployment for robotics. So reward modeling is one piece of that. On-robot reinforcement learning where you have data from deployment is also a piece of that. But that’s what I do.Michael Cho [01:22]All right, great to have both of you on. Yeah, maybe Philip you can take us through.Philip Schroeder [01:27]Yeah, yeah. Okay. So the goal of this paper was to develop a vision-language model that has strong enough reward reasoning that we can train policies from scratch through online reinforcement learning. So the model we developed is called SOLE-R1. And we show that it can serve as the sole reward signal for learning over 20 different unseen manipulation tasks through online RL. And I guess the motivating problem for this paper was that if you use just general vision-language models or VLM-based reward models that have recently come out, you get this reward hacking behavior when you try to train a policy from scratch using the predicted rewards.So here, the task is to open the drawer, and you can see that these VLM-based reward models give very high rewards just for the policy fiddling with the drawer and just slightly opening it. And then here’s another example with the Pick Cube task in RoboSuite. Again, you see that the policy is getting these very high rewards by just finding these odd camerangles and positions that trick the VLM into thinking the task is being accomplished without actually doing the task. And you can see here the SOLE-R1 predicted rewards are plotted here in blue, and the ground-truth rewards are plotted in red. And you can see that solar SOLE-R1. Much more robust to this type of reward hacking.And you kind of think of this paper as a recipe for how to train a reward model that is robust to reward hacking. And there’s kind of two major components to this that differentiate our work from prior work and really help with this reward hacking robustness. And I can quickly mention those and then go over them in more detail later. So the first is our training data mixture. We do I think three unique things with our training data mixture. One is that we include a lot of general spatial reasoning data. And a lot of this data is new. It came out with NeurIPS 2025 and ICLR 2026 papers. So at the time of this work it was data sets that were just released.So it includes just general spatial re spatial reasoning data SSR chain-of-thought And some robotics data sets, some recent one called From Seeing to Doing, and some very famous ones like Embody chain-of-thought and Robo2VLM. And we combine this with data that forces the model to reason over multiple images within the same space, capturing different moments in time. So this includes general spatial reasoning data or spatiotemporal reasoning data spot the diff and robot-specific video data that we generate and then. Also curate from Open X-Embodiment. And our video data includes failure datand I think much more failure data than prior work because we generate a lot of it ourselves in RoboCasa. So that’s one big thing.And then another big thing is that we annotate the time each time step of the videos with dense progress values and dense chain-of-thought reasoning traces to help the model explicitly reason about progress with these failure trajectories or suboptimal trajectories. And then we mix this with these artificial failure data from Open X So that’s the training data mixture. And then another novel thing we do is have the model do this chain-of-thought reasoning at each time step of the video. And this has been shown to significantly improve language model performance and visual language model performance. So and we show ablations where it helps significantly.So the models con it at each time step is reasoning about what’s changing over time between the current and previous time step and predicting a dense progress value. Relative to the previous progress value that it predicted at the last time step. And then we do a lot of experiments that I can that I can show. But let me know if you have any questions so far. I can go over in more detail the both how we generate our training datand then how we train this chain-of-thought reasoning over video.Jiafei Duan [05:56]Can you let me yeah. Yeah, I’m gonna go ahead. Yeah, no, you go ahead.Michael Cho… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

In many fields, automated research has become increasingly plausible with the improving capabilities of frontier language models, and the field robotics may be no exception. In ENPIRE, Wenli Xiao, Jia Xie, and Tonghe Zhang build a harness framework that captures physical feedback and allows for automated robotics research, minimizing human effort while training policies that achieve a 99% success rate on challenging, dexterous tasks like organizing a pin box or fastening a zip tie.Learn more in Episode 110 of RoboPapers, with Michael Cho, Chris Paxton, and Ruijie He!AbstractAchieving dexterous robotic manipulation in the real world heavily relies on human supervision and algorithm engineering, which becomes a central bottleneck in the pursuit of general physical intelligence. Although emerging coding agents can generate code to automate algorithm search, their successes remain largely confined in digital environments. We conjecture that the missing abstraction to automate robotics research is a repeatable feedback loop for real-world policy improvement: reset the scene, execute a policy, verify the outcome, and refine the next iteration. To bridge this gap, we introduce ENPIRE, a harness framework for coding agents that instantiates this physical feedback routine with four core modules: an Environment module (EN) for automatic reset and verification, a Policy Improvement module (PI) that launches policy refinement, a Rollout module (R) to evaluate policies with one or multiple physical robots operating in parallel, and an Evolution module (E) in which coding agents analyze logs, consult literature, improve training infrastructure and algorithm code to address failure modes. This closed-loop system transforms real-world manipulation learning into a controllable optimization procedure, minimizing human effort while allowing fair ablations across training recipe and agent variants. Powered by ENPIRE, frontier coding agents can autonomously train a policy to achieve a 99% success rate on challenging, dexterous manipulation tasks, such as organizing a pin box, fastening a zip tie, and tool use, a process that further accelerates when we dispatch an agent team on a robot fleet. Our results suggest a practical and scalable path toward deploying coding agents to autonomously advancing robotics in the physical world.Learn MoreProject Page: ArXiV: https://arxiv.org/abs/2606.19980TranscriptMichael Cho (00:06.088)Hello everyone, welcome to another episode of RoboPapers. Today we’re super happy to have Wenli, Jia Xie and Tonghe. They are three of the co-authors of this paper called ENPIRE: Agentic Robot Policy Self-Improvement in the Real World. Thank you so much for coming to the pod. Maybe as a start three of you can share a little bit about yourself first and then you guys can go through the paper.Wenli Xiao (00:30.821)Okay, I’ll get started. Sorry. So my name is Wenli. I’m a final-year PhD at CMU, advised by Professor Guanya Shi. Previously I did two-year internship at NVIDIA GEAR Lab with Jim Fan and Yuke Zhu. And later I joined Physical Intelligence as intern.Yeah, I did a bunch of VLA training, post-training, real-world RL and agentic stuff and now I’m mainly doing pretraining, studying the science of training and data. Yeah.Michael Cho (01:14.488)Yeah, I think just yeah.Jia Xie (01:16.463)Okay, sure. Yeah. Hi everyone. My name is Jia and I’m a master’s student at CMU, advised by Guanya. Of course we’re labmates. And I’ve interned at xDof as a research intern over the summer. And I’ve been working on some agent stuff and the extent stuff lately. So yeah, thanks for having us.Tonghe Zhang (01:43.739)Okay. Hi, my name is Tonghe. I’m also a master’s student at Professor Guanya Shi’s lab. I was doing internship at Amazon this summer and previously I’ve been working on reinforcement learning. And thank you guys for inviting us.Michael Cho (02:01.15)Alright, perfect. Great to have you guys. Yeah, Wenli, I know you have a slide, maybe you can go through with us.Wenli Xiao (02:08.228)Okay, yeah. So yeah, today the main topic is to introduce our paper, but as we all know, recently there’s a major breakthrough in AGI and that affects robotics, that is GPT-6 Astra. There are a lot of demos from Twitter that shows people apply this model to do different tasks.That related to robotics, like real-to-sim, sim-to-real, in-context learning, and code-as-policy. And we are very excited and we are kind of a little bit frustrating and also wondering how well robotics been solved. But we will leave the discussion to later section. And because even before the GPT-6 AstraAlready witnessed a huge improvement of coding agent this year because we felt like coding agent is a biggest thing in AI this year because starting from February or last year’s December we see like a lot of a lot very good coding agent like Opus and Codex. They can do a lot of stuff like starting from front-end toBack-end and build a lot of great applications. And actually, people have been trying to use these models in robotics every time there’s a breakthrough in AI. And for this year there’s a CaP-X project, which we apply the and benchmark the state-of-the-art LLM model and VLM model toTry to solve and hill-climb on the robotic simulation benchmarks and it did pretty well in simulation. And we also tried to use coding agent to just solve the very challenging manipulation benchmark. This is a very classical task which would you need to manipulate a dot to push a T-shaped object until you perfectly matches the target.Wenli Xiao (04:31.492)And this was really challenging and people find a solution or people didn’t find a solution until diffusion policy. But now we just prompt the coding agent out of the box. This is actually Codex. We prompted to write a heuristic policy with no neural network training to achieve a hundred percent success rate. It just try and try and eventually it can likeWrite very good heuristic b policy. And this applies to different coding agents like Codex Claude Code and Kimi from Moonshot. This was very exciting to me. And actually during the past two years there are many impressive works about applying coding agent for robotics.So here I just like briefly divide them into three categories, like VLA plus harness, so you can achieve great performance that can be better than either VLA or code-as-policy itself. And there’s of course Code-as-Policy, which people apply coding agent to try and practice in simulation so they can learn some skills and later apply to the real-world.Real world robotics. And there’s also physical autoresearch, which is the ENPIRE work that we are gonna discuss today. So the major difference between the first two categories and the physical autoresearch is even all of them applied the method in the real-world. For the first two part of first two categories, they need tons of human effort.So if we summarize the procedure of the researchers do these projects, it’s like they apply coding agent to write code and do some practice in simulation and also with some guidance of human, it develops some impressive skills. And then the researchers, PhD students, deploy the policy in the real-world and watch the results.Jia Xie (06:54.974)Okay.Wenli Xiao (06:55.682)And maybe for the first trial it doesn’t work, or even the robot just do nonsense. And then the student just describes what happened and type in the terminal, and then the coding agent fix it and do another trial error and apply and suddenly it works, and then can submit paper. So in this process we observe that humans are a major bottleneck.Because we become the agent’s actuators and sensors. Yeah, because the observation was from like six months ago, now with GPT 6 extra, maybe we don’t need to be the sensors of the robot. GPT can like if you set up a webcam in a third-person view, maybe GPT can observe that the policy doesn’t work.But six months ago we didn’t have such good GPT and we have GPT 5.5 and we still… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

What happens if you connect a hundred AI-powered robots to the internet and just let people give them any instructions they want? Enigma founder Jonathan Jacobi joins us to talk about his experience deploying a fleet of robots on the internet, keeping them operational, and watching people truly test the generality and robustness of frontier robotics models.Watch Episode 109 of RoboPapers with Chris Paxton and Ruijie He today!Learn MoreOriginal Post on XRobots.online website (no longer active)Enigma.incTranscriptTranscript is automatically generated by Riverside and edited by ChatGPT.Chris Paxton (00:05)Hey everyone, and welcome to another episode of RoboPapers. Today we’re super excited to have Jonathan Jacobi from Enigma here to tell us about their work in foundation models and about putting a hundred robots online.So, could you introduce yourself?Jonathan Jacobi (00:19)Yeah. Thank you, Chris. Thank you, RJ. It’s a pleasure to be on RoboPapers. Big fans.A bit about Enigma: we’re a company working on foundation models for robotics, as Chris said, as well as building innovative new interfaces for robotics. How should we interact with robots? How should the interface feel to users?We’re also interested in the back-and-forth between interface work and research on the models.About a month and a half ago, we put a hundred real robots online. You could actually interact with them over the internet. They were running our models through a variety of interfaces, and you could simply ask them to do different things.So we’re here to talk about the models, the interfaces, and the experience we’ve built.Chris Paxton (01:21)Awesome. So what did people do with all these robots that you put online?Jonathan Jacobi (01:25)The basic concept was that we wanted people to feel robotics today.We all see a lot of demos. We see a lot of cherry-picked demos—some of them are cherry-picked, some of them aren’t—and we wanted people to feel that something is here right now.One of the problems is that not enough people have robots in their houses. So how do we make enough people feel that robotics is happening now, while we still have this challenge of not having enough robots deployed in the world for the average person to use?We thought quite hard about how to make that possible. One of the best ideas we came up with was: maybe we can set up a huge warehouse with a lot of robots, put our models on those robots, and let anyone online interact with them.We also got a very cool domain: robots.online.We tried to create enough different setups that one robot could be in one scenario, another robot could be in another scenario, and people would have enough flexibility to do random things that they could really feel the AI work we’ve done.They could ask the robot to do things themselves, without the logistical challenge of putting a hundred humanoids all over the world.That’s why we used tabletop robots—arms that we customized and designed ourselves, building on public open-source designs—and let people interact with them freely.Some robots were artists. People could paint. People could do pantomime. People did very unexpected things.Someone asked one of the robots holding a paintbrush to pretend that the brush was a broom.Jonathan Jacobi (03:50)So it actually swept the floor with the paintbrush.We have some very cool examples of that that we can walk through later.The concept was that, at Enigma, we focus on generalization—text, video, audio—and we have very different interfaces and input methods for our models. We wanted people to be able to experience that generalization directly and see that it’s working today.Some of the robots were artists. Some were doing chemistry experiments. We had one that could defuse a bomb. We had one that could do lightsaber-esque fighting.Not actual lightsabers. Don’t sue me.But basically, we let people do different things with different robots online.RJ He (05:01)Okay, cool. Can you describe a little more of the interface that was set up?It sounds like each one is a single manipulator fixed to a tabletop, and then the interface is a regular chat-prompt type of thing, as if you were speaking to an AI agent. Is that right?Jonathan Jacobi (05:21)Yeah. That’s a good question.We had a static manipulator. One of them was based on the SO-101 design. We took the SO-101 and said, “We want this to look really polished.”So we brought in incredibly talented industrial designers and worked with them for about a month or two. They were working full-time on designing this new version.The robot you’re seeing on the screen is an SO-101 that we customized. The colors, the enclosure, everything was custom-designed. You can see that even the cables are hidden within the robot rather than hanging around it.These were static manipulators, and each table had a different type of interface.The artist robot had a chat interface similar to an AI agent. But we wanted to make it less boring than simply asking the robot to paint or perform pantomime, so we turned it into a game.You could give it an idea and it would paint something associated with that idea, and then you had to guess what it painted.For pantomime, you could tell it to pretend that an object was something else, and the robot would mime with it.So that was one chat-based interface.Then we had other interfaces. Let me walk through some of the things we built.This is our company page. You can see more than a hundred robots here. This is part of the actual warehouse. Those are spare robots and leftover components that we built.Here’s one of the chat interfaces where people were doing pantomime. You could do drawing, pantomime, and a few other things.And here’s another interesting interface.You could tap objects directly on the screen and then ask the robot to do something with them. You could tap the green bottle and say, “Pick this up,” then tap the orange flask and say, “Pour this into this.”Jonathan Jacobi (07:45)You could say, “Throw this here,” or whatever you wanted.We wanted to introduce a new kind of interactive interface—not just text, not just teleoperation, not just audio, but something new.Our vision is that if you extrapolate this ten or a hundred steps forward, the interface almost feels like a game. It feels like you’re playing the physical world.Think about something like StarCraft. Imagine having an interface where you can control a hundred robots. You can tap, type, drag and drop, and build agents on top of that to automate entire processes.That’s where we imagine the world going. Controlling robots becomes incredibly intuitive: you can type, tap, drag, drop, maybe talk to them.Another interface was the bomb-defusing robot. You were on a mission to defuse a bomb, and while you’re trying to do it, an operator is shouting at you: “You have to defuse it faster!” So you’re under pressure while talking to and controlling the robot.People did very unexpected things with this.Someone said, “Knock on the bomb twice at the bottom.”Someone else said, “Pour the orange flask over the pink button.”The pink button wasn’t standing upright—it was lying on the table.Chris Paxton (09:45)Can I ask—what simulation is that? That actually looks pretty good. The water physics there.Jonathan Jacobi (09:50)This is the physical world. This is not simulation.Chris Paxton (09:52)That’s not simulation? Okay. Well, that makes sense.RJ He (09:54)No way.Jonathan Jacobi (09:56)Yes. Every one of these is the physical world. None of this is simulation.Chris Paxton (10:01)Right. I knew the others were real. I don’t know why that one looked like a simulation to me. Sorry for doubting you there.RJ He (10:05)Yeah, agreed.Actually, maybe that’s a good question.First, what was the main motivation for doing this whole thing? Was it about the human-robot interface, or was it about seeing what kinds of prompts and questions people would ask?And then the second question is about this simulation-versus-real-world issue. Why did you need to do i… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Pretrained vision-language-action models for robots often fail when a task is performed in a slightly different environment. With Vision-Language Steering, you can get better generalization performance out of a frozen robot policy, by steering the sampling process of your pretrained diffusion or flow-matching policy to handle out-of-distribution changes like obstacle position or changing visual appearance. Shuo Lin and Ishneet Sukhviner Singh join us to talk about how to better generalize robot policies with no new training data.Watch Episode 108 of RoboPapers today, with Jiafei Duan and Ruijie He!AbstractWhy do pretrained diffusion or flow-matching policies fail when the same task is performed near an obstacle, on a shifted support surface, or amid mild clutter? Such failures rarely reflect missing motor skills; instead, they expose a limitation of imitation learning under train–test shifts, where action generation is tightly coupled to training-specific spatial configurations and task specifications. Retraining or fine-tuning to address these failures is costly and conceptually misaligned, as the required behaviors already exist but cannot be selectively adapted at test time.We propose Vision–Language Steering (VLS), a training-free framework for inference-time adaptation of frozen generative robot policies. VLS treats adaptation as an inference-time control problem, steering the sampling process of a pretrained diffusion or flow-matching policy in response to out-of-distribution observation–language inputs without modifying policy parameters. By leveraging vision–language models to synthesize trajectory-differentiable reward functions, VLS guides denoising toward action trajectories that satisfy test-time spatial and task requirements.Learn MoreProject Page: https://vision-language-steering.github.io/webpage/ArXiV: https://arxiv.org/pdf/2602.03973TranscriptTranscripts are lightly edited by ChatGPT.RJ He (00:01)Okay. Hi everyone. Welcome to another episode of Robot Papers. Today we are super excited to have with us Shuo Liu and Ishneet, who will be sharing their paper on Vision-Language Steering: Steering Pre-Trained Robot Policies via Vision-Language Models. Shuo, Ishneet, do you all want to introduce yourselves?Shuo Liu (00:29)Yes. Hi everyone. My name is Shuo Liu. For now I’m a research engineer at Cortex Robot AI and previously I graduated from University of Washington, where I was also a student researcher at Allen Institute for AI, advised by Professor Ranji Krishna and Professor Jiafei Duan, focus on robot learning from develop foundation models to deploying them into real world.RJ He (01:01)Isn’t it?Ishneet (01:02)Yeah. Hi everyone, my name is Ishneet. I work with Professor Jiafei on this paper, VLS, with Shuo. I’m currently at Menlo Research, where I also work on Asimov, and I’ll be going to Oxford for my undergrad later this year.RJ He (01:17)Super exciting. Amazing that you’re already having publications at this young age. It’s awesome. Okay, without further ado, Shuo, why don’t you take us through the paper?Shuo Liu (01:26)Okay, sure, yeah. Let me share the screen. Okay, so just make sure it’s it’s good to say right.RJ He (01:38)Yeah, it’s good.Shuo Liu (01:39)Yeah. Hi everyone. Today I want to share our work vision language theory. This work is finished when I was a master student at the University of Washington and advised by Professor Ranjik Krishna and Professor Jiafei Duan. And so today I want to introduce motivation of this work and the method we solve this problem and the experiment results and also some future works maybe could be done from this direction. So the problem is from the observation of some small front twining of the Like models like π0 or π0 Po, we found that current generative robot policies are can’t are kind of tightly coupled of the training specific configurations, so which cause some failing to decouple the reusable motor skills from the normal task constraints. Like for example, here we have some cubes on the table and if you want to stack them up to a tower with specific numbers you may call the π0.5 pre-trained value a policy to stack some cubes like if you want to stack two cubes and using the termination signal like the pushing the red button on the table it can do very well because you Because for this specific task you can fine twin the base policy a bit, like stack two cubes and push the button or stack four cubes to push the button. And but what if at test time you ask the base policy to stack three cubes? It’s never it’s never contained in the fine-tuning data set, so the base policyShuo Liu (03:33)Can learn how to stack two cubes or stack four cubes and push the button, but it will never know how to st how to stack three cubes and push the button. This is because the training configuration we combine the language instruction and the image observation with the task two cubes and four cubes together and the polic the base policy is never contained with task specific instruction stack three cubes. So the motor skills like moving the gripper to the cube and stack it on another one and going push the red button is already included in the base policy. But at test time if you given another out of distribution scenario or out of distribution language instruction, the base policy can always fail. So we think this is from the standard imitation learning reset. So for now if we have a base policy like take π0.5 or Momo Act 2 for example, the base policy can be trained on the behavior cloning target, which means given an expert data set which contains observation, action, and the language for each index and the optimization target is to maximize the likelihood of the base policy to be to predict the next action chunk based on current observation and the absurd language instruction. So the objective is stand static and also is very distribution dependent, which means after fun training your base policy can only capture the special and semantic correlations on the training manifold if you change the observation or language at test time the base policy may confuse. This observation is clear at our daily work because if you take one pre-trained VLA policy you just don’t want to like use it in theShuo Liu (05:53)In distribution tasks, you may change the scenario, you may change the language to the VLA. So for example, sometimes we may face observation out of distribution signal which means we may made some visual or special shift and some semantic shift which means the change of language distribution. So at that time the world may drift away from the training distribution and how we can solve it. We for now we have some two solutions. The first one is we recollect a small dataset out of distribution scenario and run another training cycle. But we need to repeat this circle for every new deployment and make the base policy useful for this specific scenario. But this is kind of misaligned with the target of we train very large VLA policy. That’s because the target is to make the base policy general generalizable but and we can using the pre-train and post-train recipe to make the skills like pick and place or open close this kind of low level skills already inside the VLA policy but we are paying the cost to relear the existing skills at test time s simply because we are not selectively adapt to the base policy at test time. So in this work we want to train we want to try another way that is steering the base policy at inference time, which means we’re taking the we’re taking the ability of generative models like photomatching or diffusion models which areShuo Liu (07:46)Broadly used in current VLAs like the action part action expert part. We can take the advan advantag advancement of their ability to denoise iteratively. So for example the diffusion model is denoising to the target distribution from Gaussian noise and using many diff denoising steps and the flow matching are learning denoising paths to from the Ga… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

Achieving human-like dexterity is the next frontier for robotics, and yet dexterity data is often subtly hard to scale. Real-world dexterity data, including things like finger-pose estimates, is often slightly off, making it physically invalid and hard to execute on real hardware and hard to learn from.DO AS I DO is an algorithm for reconstructing and retargeting monocular RGB videos to robot hands, outperforming the state of the art and even working from generated videos. Bhawna Paliwal, Haritheja Etukuru, Will Liang, and Mahi Shafiullah join us to tell us more.Watch Episode 107 of RoboPapers, with Chris Paxton and Jiafei Duan, now!AbstractHow can we scalably generate data for robotic manipulation, especially on human-like platforms such as dexterous multi-fingered hands? Learning from human videos has recently emerged as a likely answer to this question. However, difficulties in estimating hand-object interaction and crossing the human-to-robot embodiment gap have hindered the adoption of abundant monocular RGB-only human videos as the primary source of robot manipulation data. In this work, we present DO AS I DO, an algorithm to reconstruct and retarget monocular RGB human videos to multi-fingered dexterous robotic hands. DO AS I DO reconstructs hand-object interactions from various egocentric and exocentric in-the-wild video sources. The algorithm then retargets these hand-object interaction estimates into a sequence of actions executable in the real world, yielding robot-complete manipulation data from disparate human videos. Overall, DO AS I DO outperforms previous state of the art in estimating hand-object interactions and extracting dexterous manipulation trajectories from RGB videos, as we show in experiments on datasets with ground truths and on a dataset of video clips collected online. Our experiments enable us to propose an efficacy playbook for practitioners collecting human data for manipulation.Learn MoreProject Page: https://do-as-i-do.com/ArXiV: https://arxiv.org/abs/2606.19333Github: https://github.com/malik-group/do-as-i-doOriginal thread on XTranscriptLightly edited for transcription errors, names, technical terminology, punctuation, and readability. Conversational phrasing is otherwise preserved. Editing done automatically with ChatGPT.Chris Paxton (00:05.479)Hey everyone, welcome to another episode of RoboPapers. Today we have some great work from Bhawna, Haritheja, Will, and Mahi: “Do as I Do: Dexterous Manipulation Data from Everyday Human Videos.” Welcome to the show, guys. Could you introduce yourselves quickly before we get started?Bhawna Paliwal (00:24.977)Yeah, hi, I’m Bhawna. I’m a PhD student here at UC Berkeley, advised by Jitendra Malik.Haritheja Etukuru (00:32.692)I’m Haritheja. I’m a second-year PhD student at Berkeley, also advised by Jitendra.William Liang (00:37.929)I’m Will. I’m also a second-year PhD student, advised by Jitendra and Pieter.Mahi Shafiullah (00:42.606)Hey everybody, I’m Mahi. I’m a postdoc currently at Berkeley, and I’m also spending some time at Amazon FAR right now.Chris Paxton (00:50.983)Awesome. Well, yeah, would you guys like to tell us a little bit about this? What have you guys been working on?Bhawna Paliwal (01:01.499)So I’ll just start with a little introduction to the paper. So mainly the motivation behind this work was how can we utilize the massive video data available on internet for training robots. And most of the prior work has been which which has involved learning from human videos, has involved co-training with robot data, which which eventually involves having like hundreds or thousands of hours of teleop data combined with, let’s say, twenty thousand hours of video. So our objective in this paper is to understand. Whether there is some intermediate representation, let’s say in in 3D, where we can extract the right kind of information from human videos and make learning from human videos much more efficient. So essentially every human video where hand-object interaction is going on turns into a separate trajectory on which you can train like a robot model. So prior work has looked into this problem mostly through by collecting their own data and maybe having like specific settings where they’re They have like specific angles and specific robot specific hands hand-object tasks available. But here we look at internet videos in general where there might be a lot of occlusion, motion blur and so on, which are not necessarily recorded for robot learning. Yeah.Chris Paxton (02:21.331)Right. Some of the videos here in the highlights look incredibly good, particularly with the dexterous hand. So how does this work?Bhawna Paliwal (02:35.113)So I think As like broadly there are like two steps. First we have reconstruction where we reconstruct the object in 3D, as shown here. And here and then we retarget it onto a robot hand, in our case the Sharpa hand. And for reconstruction, like the hardest part was object tracking. That is most of the times when we are looking into hand-object videos, the objects are mostly occluded, especially the tiny objects such as this one. There are like just fewer than a hundred pixels where the object is actually visible. And reconstructing the object becomes hard first. And secondly, tracking the object becomes even harder, even if you know the shape. So, in this work, we first we utilize like all the computer vision progress that has been made recently, especially models like SAM 3D, which can just like take a single image and give you the shape and pose of the object in that particular frame. And then we come up with our own method of how do you utilize the SAM 3D kind of methods to also track the object over time. And the motivation behind was just behind that was just this that SAM 3D has been trained on lots of images and artificially constructed images where like some particular object is blurred out or it has been hidden by some other object. And then so it knows when the how the object looks when it’s like heavily occluded. And then what we do is like we predict the shape and pose at one particular frame and then propagate it to next into like further frames. while keeping the constraint that the shape of the object should not change too much and things like that. So I let Will explain the retargeting part. Yeah. SoJiafei Duan (04:18.402)Yeah, I think I have a sorry. I have a question. Yeah, so I actually I read I read the paper quite a few times. and I find it very interesting. So what we are seeing here is an open-loop replay, is that my understanding? Right. So would that be like any limitation that you guys foresee? I mean, maybe I’m jumping ahead of time, but I’m just curious, right? Like for example, certain tasks when you are, you know, doing requires some level of precision or perception, then how will you address that? Or is it just more of the motion that matters at this point?Bhawna Paliwal (04:51.655)Yeah, so these trajectories are right now open loop and the how we foresee this should be changed into a perceptive task, which is definitely necessary, is by training taking this reference, domain randomizing the object position, and then training an RL policy, let’s say using point cloud as input. So this will involve just like since you already have the reference trajectory in simulation, you also know the object shape, and then you can like train an RL policy. The object pose is changing and let’s say like even the your environment is changing, lighting and so on, and then you can do all these kind of domain randomization tricks to kind of make it work for one object. And then once you have lots of such policies, you can probably distill it into one big model with let’s say behavior cloning. yeah. Mahi, maybe you would like to add more things.Mahi Shafiullah (05:46.155)Yeah, I mean I think one of the things that I think about when it comes to this project is really that we’re exploring how to get robot data from kind of arbitrary i… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com

World Action Models are becoming more popular in robotics, as they learn to predict the world jointly with learning how to act on it. However, these world predictions are usually purely based on reconstructing color images from video. This is limiting, because color is far from the most important quality for a robot moving around in the world — more important are qualities like 3D geometry and object semantics. In Flex-π, Ge Yan, Jesse Zhang, and team train a 6 billion parameter world action model to do exactly this, predicting 3D pointmaps and DINO features along with color. This results in a policy which is much more demonstration-efficient, generalizes well, and can perform complex long-horizon tasks.Learn more in Episode 106 of RoboPapers, hosted by Jiafei Duan and Ruijie He!AbstractWorld-action models (WAMs) predict the future to act better, but nearly all of them predict only RGB latents, trained purely for pixel reconstruction, with no explicit signal for the 3D geometry or object semantics manipulation needs. We find a surprising free lunch: the same frozen video-generation VAE that encodes RGB also encodes 3D pointmaps almost losslessly, with no pointmap-specific training at all. This lets us supervise Flex-π, a 6B-parameter WAM, on 3D geometry and object-centric DINO semantics alongside RGB, at no cost in new sensors, new pre-training, or inference latency. Every visual signal is projected into this shared latent space and denoised jointly with actions inside a Mixture-of-Transformers backbone; per-stream dropout with cross-modality forcing then lets a single trained checkpoint run on any subset of these streams, from a fast action-only mode to full joint generation. The result is a policy that is exceptionally demonstration-efficient and generalizes well, beating the strongest baselines by up to 2-7× on dexterous, precise, real-world bimanual manipulation tasks both in and out of distribution, all while running faster than π0.5. Our project website: this https URLLearn MoreProject Page: https://flex-pi.github.io/ArXiV: https://flex-pi.github.io/Thread on XTranscriptTranscript edited and prepared with automated transcription and checked by GPT 6. There may still be errors.Jiafei Duan (00:04.231)Okay, sure, sure, sure. Yeah, hello everyone. Welcome to a new episode of RoboPapers. We passed our hundredth paper mark and this is, I think, 102 or 103. And and today we have like two of the returning speakers from the previous episodes, Jesse and Ge Yan, to come on and really share about their new work, Flex-π. So maybe do you guys want to Reintroduce yourself a bit more and then I think for some of the audience who never heard the previous podcast before. Yeah.Ge Yan (00:38.151)Yeah, yeah, hi Jiafei, and thanks for having us. So I’m right now like a second year PhD student who is going to enter my third year at the University of Washington. And my PhD advisor is Dieter Fox. And now I’m working also closely with TRI and Jesse. Yeah, Jesse, it’s yours.Jesse Zhang (01:01.599)Yeah, I’m a postdoc at UW and I’m advised by Dieter Fox, and Abhishek. And so I started this Ge and I worked on this project for the past how long has it been? Maybe almost a year, yeah, it’s been a while. And it’s changed quite a bit since the initial version was definitely not a WAM. so it’s been very interesting how this has turned out.Ge Yan (01:18.833)Yeah, it’s been a long time.Ge Yan (01:26.567)Yeah, great.Jiafei Duan (01:27.145)Yeah, it’s it’s been it’s been a quite a while. I mean, I remember I was still there when all the people collecting data.Ge Yan (01:35.464)Yeah.Jesse Zhang (01:35.623)Yeah, yeah. I mean, even like Phil presented on on ideas that led to this and they were very different than what it is now.Jiafei Duan (01:42.677)Yeah, yeah. Yeah. I mean I mean for audience like just context, we are from the same group, so I I know exactly what this is about. But yeah, it definitely changed. But yeah, we hope to hear from Ge exactly what Flex-π is all about. So yeah, Ge, you want to take over.Ge Yan (01:49.629)Yeah.Ge Yan (02:00.219)Okay, thank you. Okay, hello everyone. Today we’re going to introduce Flex-π, a multi-stream world-action model with compute flexibility. I think we framed it as a policy that can do one checkpoint, any observed inputs, any generated features, and the speed accuracy operating point you pick and deployment. And this work is done by me, Jinghao, Yuzhi, Lori, Minwen, Jesse, and Dieter at the University of Washington. So I think recently we’ve seen a lot of work on on WAMs.I think it starts from the trend of very good video generation model, and people start to leverage those video priors from those very large video models and use it for action prediction. And we’ve seen a lot of progress, including like Fast-WAM, LingBot-VA, DINO most recent DINOTube paper work and and a lot of others. And I think we know that the WAMs are great, but they usually only predict RGB latent and that is specifically used to for reconstruction. And the previous work like DINO or DreamZero, so they really shows us that the world modeling objective itself enables better action prediction.And I think it is quite intuitive right now, like with with so many results show its benefits. So I think in our work we kind of want to ask a question like why not predict additional and that is especially manipulation-centric visual signals. So we really want to leverage those other signals because we want to further push the world modeling objective strengths, and we want to do some very high precision and dexterous tasks, and we want to achieve better demo efficiency and generalization. So the right is a a multi-view.Ge Yan (04:02.367)RGB video shown in a YAM setup that a model do this Soft-Bag Zipping task. And so right now you can see it’s purely RGB right? And so for Flex-π we focus on like two additional signals. So for better Demo efficiency and the generalization. So one is DINO. So we know the DINO had provides very good object-centric semantics that can be very good for generalization, especially in those cloud cloud environments. And also the 3D point. I think this 3D information really provides a lot of demo efficiency, lots of model to achieve sometimes even have high performance with much less data.As as we know, there’s no like free lunch, right? as we introduce like additional signals, it can like possibly like always introduce like additional cost. So that that really comes to one problem, because the most WAMs the the inference mode is decided at training time. So like what kind of inputs the model takes and what kind of features the model needs to generate. And we know that predicting the future makes us better, but actually w when you generate those at the inference time make it pretty slow. And it’s usually we see a lot of model that joint and generate those features, it’s very hard to do real time.And especially with our additional visual signals, that makes even makes it even harder. So we we kind of like want to do like one model that can do both. they can do like action-only prediction, and then you can also predict the futures as other WAMs. And it can be decided at the deployment time.Ge Yan (06:00.741)And before introducing our exact architecture, we want to share that because we build on top of the Wan video model backbone. So this the bottom video shows after training on internet-scale videos, the Wan model, this backbone has a very good physical prior. You can see it can generate this looks reasonable like videos. Although there you can see there are still some artifacts, but it really sh it shows the model can learn like the r what waterwheel looks like a reasonable behavior, even though it’s not exactly the the right. And and one thing about the one backbone is like the VAE part. So VAE can be regarded as an encoder.So for a VLA you can have like any those image language encoder, and for these one models, we use something called a Wan 2.2 VAE. … This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit robopapers.substack.com