RoboPapers

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 datasets that have different action distributions. There is interesting ongoing work on making synthetic data match human action distributions, but that wasn’t something we focused on here.Rose Hendrix: And just to be clear: all of our policies are trained strictly on synthetic data. There’s no human data, teleoperation data, or real-world robot data involved in their training.Abhay Deshpande: Altogether, we have 1.7 million demonstrations across more than 94,000 environments, 11,000 object assets, and 9,000 receptacles, totaling around 5,700 hours of data. That makes it one of the largest open robotic datasets available, and the nice thing is that you can scale compute to generate it.In terms of hours, that puts us well ahead of datasets like DROID, Open X-Embodiment, and RoboCasa365. Where we really stand out is environment diversity. By building on MolmoSpaces and its indoor environments, we get more than 94,000 unique environments—far more than most other datasets provide.The policy architectures and evaluationsMaya Guru: We trained three different model classes. All take visual information and proprioception as inputs: our flagship MolmoBot model, MolmoBot-SPOC, and MolmoBot-Pi0. All three achieve zero-shot transfer to the real world and generalize to new environments, though at different performance levels.For MolmoBot and MolmoBot-SPOC, the inputs include external and wrist cameras for the Franka, or head and wrist cameras for the RB-Y1, along with a language task instruction, proprioception, and optional point conditioning. For example, if we want the robot to open a specific drawer, we can point to that drawer and instruct it to open it.Our flagship MolmoBot architecture uses the Molmo2 vision-language model with an action expert. MolmoBot-SPOC is adapted from previous work and uses a lightweight transformer. One reason to offer that architecture is that it makes reinforcement-learning fine-tuning more accessible.We also trained MolmoBot-Pi0, using the π₀ archi

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 trade-off. If you want to detect slip faster, you may make more mistakes and trigger more false alarms. Prior work rarely evaluates that trade-off directly, and that’s where we started.Before going into the details, let’s look at some demos. Here, SlipSense is running in real time on UMI parallel-jaw grippers. The orange component on the fingertip in all three videos is our TacV5 sensor. It’s a multimodal tactile sensor developed by our team. It measures normal force using a piezoresistive array, as well as high-frequency vibrations using an accelerometer sampled at 8 kHz.This sensor wasn’t originally designed for slip detection. It’s a general-purpose tactile sensor, but we found that it works very well for this task. In all three videos, you can see the model’s predictions updating live. In the left video, when the operator pulls on the object, SlipSense detects the slip almost immediately. In the middle video, it can also detect rotational slip.Michael Cho: I’m curious: physically, where exactly is the three-axis accelerometer? You have this TacV5 device with a 32-by-32 array of taxels. Is the accelerometer at a particular taxel, or somewhere else? I imagine the position would affect the vibrations it senses.Tao Yu: The module is really like a fingertip. The sensor array is on the surface, but inside there’s a PCB that reads out the sensor, and the accelerometer is on that same PCB.Michael Cho: I see. So somewhere inside the module.Tao Yu: Exactly, and rigidly coupled to the surface.Michael Cho: When you’re measuring vibration, you’re using all three axes of the accelerometer?Tong Jian: Yes. We use X, Y, and Z, and we also try to compensate for gravity and motion-induced acceleration so that we can focus on friction-induced vibrations, which are the slip features we’re interested in.Tao Yu: Just to add some background on why we did this research: a lot of the questions we get when talking about our piezoresistive tactile sensor are about its limitations. One draw

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 [05:56]Yeah, let’s go over the training data. Sorry, Jiafei. No, I was gonnask about the training data. I mean it’s a nontrivial amount of training data. It’s curious how you get that, like the dense annotation. Jiafei whether that’s your question.Philip Schroeder [06:10]Yeah, so we w most of the novel data we’re generating is in sim. So we’re leveraging the ground-truth simulator state at each moment. And I have some slides for sh walking through that and I can I can show that next. Jiafei if you have a question, I can answer that before I show that.Jiafei Duan [06:26]Yeah, so maybe like I’m curious about the reward hacking, right? So you showed the first example, where I guess it’s like it’s non-trivial in the sense whereby you are trying to showcase a task where it looks like it’s grasping, but when the from a different angle it might look slightly different. Is that my understanding? Okay. So is it a perception problem or like do you think it’s like the a result of perception that leads to reward hacking or you know?Philip Schroeder [06:31]Yeah. Yes. Yeah, so itJiafei Duan [06:55]or what do you think is the you know, how do we quantify this? Yeah.Philip Schroeder [06:59]Yeah, so it’s definitely a perceptual failure. Where the policy is just finding out-of-distribution visual input by just ‘cause it has so many degrees of freedom with moving the camera into any position. And since we’re already a little bit out-of-distribution with this task setup because there’s probably not a lot of rope like the robotics data inside video data inside the training data for these visual language models. But especially when you give the policy this opportunity to optimize against the VLM’s sense of progress. And it’s always easier to just find these weird camerangles than it is to just do the task. It’s always easier to get these false positive states than to get the true positive.Jiafei Duan [07:41]I see. I see. Maybe my question would be like humans also have this, right? Like

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 wanna see how much it can do. So here is an example, we assemble the robot arm used some DaMiao motors because we are too lazy. We didn’t want to implement the firmware and we didn’t wanna implement the low-level control. SoWe just plug in the USB-CAN to the computer that has Claude Code installed and we prompt it to implement the impedance controller for the arm and try it by itself. And with great expectation and the Claude just like just think and write, think and write, and after five minutes it finished. And here’s the result.Michael Cho (08:50.424)Gosh. Ouch.Wenli Xiao (08:51.676)Yeah. And we were w glad we muted this video. Yeah, we were screaming about it. Yeah, and then we took another night to redo the 3D printing and assemble the arm again. Yeah.Chris Paxton (08:54.443)Did that break the robot? Is it? Yeah.Michael Cho (09:03.74)MyMichael Cho (09:13.212)Well, you guys were very brave, man.RJ He (09:13.527)I think you’re particularly making Chris sad, given his background.Wenli Xiao (09:16.732)Yeah. So well it’s a little bit frustrating, but because autoresearch works so well in simulation that it works and your capacity to make a Twitter post. And there’s a lost also a lot of great applications of applying autoresearch in sim and digital world. So we started to thinkMichael Cho (09:20.366)Ouch, ouch.Wenli Xiao (09:43.857)What’s different between autoresearch in sim versus in the physical world? And we think that’s because in simulation or in sandbox you can do reset for infinitely many times. You have a verifiable reward, you can heal climate. And of course there’s a sandbox, you won’t break anything. But that may not be true because maybe withA S I it can jailbreak the sandbox. But for robotics, we don’t even have this harness or sandbox in the real-world. So that’s the motivation of the entire project. So for this project we wanna take a wild step and see like what’s the minimum setup we can do to let AGI or coding agent to actually do research in the ph