正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

How do we run RL with real-time chunking (RTC)? In this work we figured out how to use a small RL policy with a large robot foundation model, where the RL policy observed more recent images (due to faster inference) and steers the policy toward better behaviors! A fun... show more

Sergey Levine

139,648 subscribers

45,721 次观看 • 18 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

4 条评论

Stephen 的头像

Stephen18 天前

"real-time chunking (RTC)" — the quiet admission vla inference is too slow for the real world, so a small rl policy rides shotgun. if async breaks the markov assumption, how much rl theory gets rewritten just to ship robots that don't fall over?

Abby 的头像

Abby17 天前

The fresher camera input is the part that caught my attention. How does this hold up when inference latency varies, rather than staying at a fixed delay? That seems like an important test for teams running perception and control on shared hardware.

JK 的头像

JK18 天前

Sergey, I've followed your work for some time now and you're a heavyweight in the field of robotics. Your utilzation of integrated sensors and specialized RL environments per task is world-class!

Noorullah Jamakzai 的头像

Noorullah Jamakzai17 天前

nice framing. once perception freshness is part of the policy, inference latency stops being an infra metric and becomes a behavior variable. feels like a very underexplored lever

相关视频

We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & Andrew Wagenmaker

We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & Andrew Wagenmaker

Sergey Levine

38,324 次观看 • 3 个月前

Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that 1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data; 2. runs RL fine-tuning online with the VLM reward model. 🔗 🧵↓

Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that 1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data; 2. runs RL fine-tuning online with the VLM reward model. 🔗 🧵↓

Jingyun Yang

88,312 次观看 • 2 年前

I've long wondered if we can make a humanoid robot do a 𝘄𝗮𝗹𝗹𝗳𝗹𝗶𝗽 - and we just made it happen by leveraging 𝗢𝗺𝗻𝗶𝗥𝗲𝘁𝗮𝗿𝗴𝗲𝘁 with BeyondMimic tracking! This came after our original OmniRetarget experiments, with only minor tweaks to RL training: relaxing a termination threshold and removing one reward term. The policy achieved a 𝟱/𝟱 success rate in our real-world experiments, showing the strength of high-quality, interaction-preserving motion retargeting combined with BeyondMimic’s minimal RL tracking. Here is the updated arXiv: (In Sec. V. A)

I've long wondered if we can make a humanoid robot do a 𝘄𝗮𝗹𝗹𝗳𝗹𝗶𝗽 - and we just made it happen by leveraging 𝗢𝗺𝗻𝗶𝗥𝗲𝘁𝗮𝗿𝗴𝗲𝘁 with BeyondMimic tracking! This came after our original OmniRetarget experiments, with only minor tweaks to RL training: relaxing a termination threshold and removing one reward term. The policy achieved a 𝟱/𝟱 success rate in our real-world experiments, showing the strength of high-quality, interaction-preserving motion retargeting combined with BeyondMimic’s minimal RL tracking. Here is the updated arXiv: (In Sec. V. A)

Zhen Wu

1,051,768 次观看 • 1 年前

Behavioral Foundation Models (BFMs) trained with RL are secretly more powerful than we think. BFM’s directly output a policy believed to be near-optimal given any reward function. Our new work shows that they can actually do much better:

Behavioral Foundation Models (BFMs) trained with RL are secretly more powerful than we think. BFM’s directly output a policy believed to be near-optimal given any reward function. Our new work shows that they can actually do much better:

Harshit Sikchi

44,479 次观看 • 1 年前

We just released the world's first open-source humanoid physical strike AI! The AI is trained with large-scale self-play RL with a pre-trained ASE model from Jason Peng. Thank ZiyiBird for this amazing work.

We just released the world's first open-source humanoid physical strike AI! The AI is trained with large-scale self-play RL with a pre-trained ASE model from Jason Peng. Thank ZiyiBird for this amazing work.

Flood Sung

101,796 次观看 • 3 年前

We developed an RL method for fine-tuning our models for precise tasks in just a few hours or even minutes. Instead of training the whole model, we add an “RL token” output to π-0.6, our latest model, which is used by a tiny actor and critic to learn quickly with RL.

We developed an RL method for fine-tuning our models for precise tasks in just a few hours or even minutes. Instead of training the whole model, we add an “RL token” output to π-0.6, our latest model, which is used by a tiny actor and critic to learn quickly with RL.

Physical Intelligence

445,939 次观看 • 6 个月前

Robots don’t learn survival the way humans do 🤖 Toru shares how teleoperation and multi-sensory policy learning can fast-track robots toward foundational human skills.She then dives into a core RL question: how do we design learning environments and reward functions that lead to useful, optimizable behaviors? Watch the full session here:

Robots don’t learn survival the way humans do 🤖 Toru shares how teleoperation and multi-sensory policy learning can fast-track robots toward foundational human skills.She then dives into a core RL question: how do we design learning environments and reward functions that lead to useful, optimizable behaviors? Watch the full session here:

Cohere Labs

20,841 次观看 • 7 个月前

🤖Adding new RL algorithms to LeRobot just got much easier. Demo: HIL-SERL training with a SAC-based RL algorithm on an SO-100 for a hole-in-hand peg-in-hole task. Sparse reward, only 30 offline demos mixed with live robot experience, and ~1 hour of online training with human interventions only when the policy fails. The bottom graph tracks intervention rate: high at the start, steadily dropping as the policy improves. The refactor separates algorithm logic from training infrastructure: • RLAlgorithm owns learning logic • RLTrainer handles orchestration • DataMixer combines rollouts, demos, interventions, and future data sources Adding an RL algorithm now looks much closer to adding a policy: one algorithm file, one config, one registry entry. SAC is first. RLT, RECAP, ConRFT, QC-FQL, DSRL, and VLA RL fine-tuning next! Thomas Wolf clem 🤗

🤖Adding new RL algorithms to LeRobot just got much easier. Demo: HIL-SERL training with a SAC-based RL algorithm on an SO-100 for a hole-in-hand peg-in-hole task. Sparse reward, only 30 offline demos mixed with live robot experience, and ~1 hour of online training with human interventions only when the policy fails. The bottom graph tracks intervention rate: high at the start, steadily dropping as the policy improves. The refactor separates algorithm logic from training infrastructure: • RLAlgorithm owns learning logic • RLTrainer handles orchestration • DataMixer combines rollouts, demos, interventions, and future data sources Adding an RL algorithm now looks much closer to adding a policy: one algorithm file, one config, one registry entry. SAC is first. RLT, RECAP, ConRFT, QC-FQL, DSRL, and VLA RL fine-tuning next! Thomas Wolf clem 🤗

LeRobot

31,237 次观看 • 5 个月前

With Harvard University, we built a ‘virtual rodent’ powered by AI to help us better understand how the brain controls movement. 🧠 With deep RL, it learned to operate a biomechanically accurate rat model - allowing us to compare real & virtual neural activity. →

With Harvard University, we built a ‘virtual rodent’ powered by AI to help us better understand how the brain controls movement. 🧠 With deep RL, it learned to operate a biomechanically accurate rat model - allowing us to compare real & virtual neural activity. →

Google DeepMind

229,082 次观看 • 2 年前

So we did a bunch of projects with real world reinforcement learning - but it was often too inefficient to be practical to train tabula rasa. This suggests we need better priors, but acquiring these from on-robot data can often be expensive as well. In our recent work, we show that despite being fundamentally inaccurate, simulation can guide provide a cheap way to guide real-world RL finetuning to be super efficient! We propose Simulation-Guided Fine-Tuning (SGFT) - a simple paradigm for sim2real finetuning that uses simulation to provide reward shaping that accelerates real world RL finetuning *beyond* just providing an initialization. TLDR: Use value functions from sim to shape rewards for real-world RL, see large sample efficiency improvements 🧵(1/6)

So we did a bunch of projects with real world reinforcement learning - but it was often too inefficient to be practical to train tabula rasa. This suggests we need better priors, but acquiring these from on-robot data can often be expensive as well. In our recent work, we show that despite being fundamentally inaccurate, simulation can guide provide a cheap way to guide real-world RL finetuning to be super efficient! We propose Simulation-Guided Fine-Tuning (SGFT) - a simple paradigm for sim2real finetuning that uses simulation to provide reward shaping that accelerates real world RL finetuning beyond just providing an initialization. TLDR: Use value functions from sim to shape rewards for real-world RL, see large sample efficiency improvements 🧵(1/6)

Abhishek Gupta

13,637 次观看 • 1 年前

I made a RL policy that guesses where a picture was taken without GPS data It continuously learns, updating its weights with every use in realtime -- over the weekend it improved 13.9% with <100 images Best of all, it does this without ever storing any image data, link below

I made a RL policy that guesses where a picture was taken without GPS data It continuously learns, updating its weights with every use in realtime -- over the weekend it improved 13.9% with <100 images Best of all, it does this without ever storing any image data, link below

Surya Dantuluri

168,785 次观看 • 1 年前

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Davide Scaramuzza

34,920 次观看 • 2 年前

The robot climbs stairs🏯, steps over stones 🪨, and runs in the wild🏞️, all in one policy, without any remote control! Our #CVPR2023 Highlight paper achieves this by using RL + a 3D Neural Volumetric Memory (NVM) trained with view synthesis!

The robot climbs stairs🏯, steps over stones 🪨, and runs in the wild🏞️, all in one policy, without any remote control! Our #CVPR2023 Highlight paper achieves this by using RL + a 3D Neural Volumetric Memory (NVM) trained with view synthesis!

Xiaolong Wang

113,062 次观看 • 3 年前

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!

Avi Chawla

138,769 次观看 • 7 个月前

If you want a robot to do something well, you need to know how to talk to it. If you don't, you can learn, with Semantic Action RL! In our paper, Jagdeep Bhatia, Andrew Wagenmaker, Will Chen show how RL over VLA prompts enables new tasks and learns blazing fast in the real world!

If you want a robot to do something well, you need to know how to talk to it. If you don't, you can learn, with Semantic Action RL! In our paper, Jagdeep Bhatia, Andrew Wagenmaker, Will Chen show how RL over VLA prompts enables new tasks and learns blazing fast in the real world!

Sergey Levine

31,707 次观看 • 3 个月前

These new hands are the perfect companion for Atlas. With 13 degrees of freedom, they are directly actuated, built for high fidelity simulation to enable sim-to-real RL, and designed from the ground up for dexterous physical work. Learn how we built robot hands for modern AI and real work:

These new hands are the perfect companion for Atlas. With 13 degrees of freedom, they are directly actuated, built for high fidelity simulation to enable sim-to-real RL, and designed from the ground up for dexterous physical work. Learn how we built robot hands for modern AI and real work:

Boston Dynamics

366,549 次观看 • 8 天前

Lecture 10 of my course! Nominally on regularization in RL, so I discuss the evolving role of the KL penalty in RL, but also a set of nice RL papers that explain what RL helps models generalize better than SFT -- with theory supporting it. When going through these, it's so interesting how seasonal problems in ML are. Lots of problems from controlling reward models overopt will rhyme as we try to control rubrics for agents. 00:00 Intro & the role of regularization 02:50 The KL penalty in RL 10:53 RL as a reverse KL loss 21:09 Why RL generalizes better than SFT 25:15 Other regularization tools Just a few videos left as I get to the end of the course. Thanks all, and keep sending questions. Spread the word if you have a second.

Lecture 10 of my course! Nominally on regularization in RL, so I discuss the evolving role of the KL penalty in RL, but also a set of nice RL papers that explain what RL helps models generalize better than SFT -- with theory supporting it. When going through these, it's so interesting how seasonal problems in ML are. Lots of problems from controlling reward models overopt will rhyme as we try to control rubrics for agents. 00:00 Intro & the role of regularization 02:50 The KL penalty in RL 10:53 RL as a reverse KL loss 21:09 Why RL generalizes better than SFT 25:15 Other regularization tools Just a few videos left as I get to the end of the course. Thanks all, and keep sending questions. Spread the word if you have a second.

Nathan Lambert

89,686 次观看 • 2 个月前

Today we give Atlas dexterous hands, with the strength to take on real work, and the simplicity for mass manufacturing. Like the body, the hands are designed with high fidelity simulation in mind, to enable sim-to-real RL. If you give dexterous hands to a robot, make them do dexterous things. More thoughts here Blog: Video:

Today we give Atlas dexterous hands, with the strength to take on real work, and the simplicity for mass manufacturing. Like the body, the hands are designed with high fidelity simulation in mind, to enable sim-to-real RL. If you give dexterous hands to a robot, make them do dexterous things. More thoughts here Blog: Video:

Alberto Rodriguez

556,312 次观看 • 8 天前

Open Duck Mini is standing up using a policy learned with RL in simulation! We are still pushing locomotion, so if you have some knowledge of Isaac Gym and want to help feel free to join us! Thomas Wolf

Open Duck Mini is standing up using a policy learned with RL in simulation! We are still pushing locomotion, so if you have some knowledge of Isaac Gym and want to help feel free to join us! Thomas Wolf

Antoine Pirrone

66,442 次观看 • 2 年前

You might have seen the WuBOT performing at the 2026 Spring Festival Gala; however, most high-dynamic extreme motions you see are executed by overfitted tracking policies. Until now, training a unified policy capable of performing various extreme motions with a high success rate remained an unsolved challenge. We spent an entire year digging into the barrier between general tracking and extreme physical behaviors. After burning through dozens of G1 robots, we finally identified the bottleneck of learning and physical executability. With these discoveries, we developed OmniXtreme: the first general policy that can execute diverse extreme motions, including consecutive flips, extreme balancing, and even breakdancing with rapid contact switches! This capability is achieved by pre-training a flow-based generative control policy and then post-training with actuation-aware residual RL for complex physical dynamics—a step we found critical for successful real-world transfer. This work is a joint collaboration with Unitree. Together, we are pushing the physical limits of humanoid robots. It is incredibly exciting to see a general "robot gymnast" and "robot breakdancer" come to life! It was also our first time publishing a paper with XingXing, which was an enlightening experience. The model checkpoints are now released—we welcome you to play with them! 📦 📄 Paper: 🌐 Project: 💻 Code:

You might have seen the WuBOT performing at the 2026 Spring Festival Gala; however, most high-dynamic extreme motions you see are executed by overfitted tracking policies. Until now, training a unified policy capable of performing various extreme motions with a high success rate remained an unsolved challenge. We spent an entire year digging into the barrier between general tracking and extreme physical behaviors. After burning through dozens of G1 robots, we finally identified the bottleneck of learning and physical executability. With these discoveries, we developed OmniXtreme: the first general policy that can execute diverse extreme motions, including consecutive flips, extreme balancing, and even breakdancing with rapid contact switches! This capability is achieved by pre-training a flow-based generative control policy and then post-training with actuation-aware residual RL for complex physical dynamics—a step we found critical for successful real-world transfer. This work is a joint collaboration with Unitree. Together, we are pushing the physical limits of humanoid robots. It is incredibly exciting to see a general "robot gymnast" and "robot breakdancer" come to life! It was also our first time publishing a paper with XingXing, which was an enlightening experience. The model checkpoints are now released—we welcome you to play with them! 📦 📄 Paper: 🌐 Project: 💻 Code:

Siyuan Huang

108,684 次观看 • 7 个月前