Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Introducing RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning. 7 real robot tasks, 900/900 successes. Up to 250 consecutive trials in one task, running 2 hours nonstop without failure. High success rate against physical disturbances, zero-shot, and few-shot adaptation Our first step toward a deployable robot learning system.

Kun Lei

1,446 subscribers

92,064 views • 11 months ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

30 Comments

Kun Lei11 months ago

If you're troubled by insomnia, you can come and count for this demo.🤣🤣🤣 Task Timeline: Folding: 05:18-16:09 Juicing-placing: 16:09-19:09 Juicing-removal: 19:09-21:21 Pouring: 21:21-23:12 Unscrewing: 23:12-24:57 Bowling: 24:57-27:26 Push-T: 27:26-30:21 OverallJuicing: 30:21-31:57

Kun Lei11 months ago

Human vs. Robots: Guess who will win?

Kun Lei11 months ago

RL-100 introduces a three-stage pipeline: (i) Diffusion policy to acquire human priors; (ii) Iterative offline RL, PPO-style objective applied in the denoising process; to deliver conservative, near-monotonic improvements; (iii) Online RL to eliminate residual failure modes.

Kun Lei11 months ago

Steer policy to 100% success after 210 trials.

Kun Lei11 months ago

Takeaway: 1) Variance clipping is valid for stable exploration - variance clipping in the stochastic DDIM sampling process. 2) Epsilon prediction is more suitable for RL: large noise schedule for exploration 3) Reconstruction is crucial for visual robotic manipulation RL as it mitigates representational drift and improves sample efficiency. 4) On a relatively clean scene, the 3D variant learns faster and attains a higher final success rate. 5) CM effectively compresses the iterative denoising process without sacrificing control quality, enabling high-frequency deployment.

Kun Lei11 months ago

Execution Efficiency: 1) CM (RL) > DDIM (RL) > DP3 (IL) > DP (IL); 2) RL > Human teleoperation

Kun Lei11 months ago

High-frequency control is crucial for real robot deployment. Achieving a policy inference frequency of about 378Hz.

Kun Lei11 months ago

A huge thank you to my incredible team for all the hard work, dedication, and teamwork. RL-100 team: @kunlei15; @imhuanyuli; @manutd_moon; Zhenyu Wei; @Lingxiao234; @Zhennanjia13007; Ziyu Wang; Shiyu Liang; @HarryXu12 Paper link:

Kun Lei11 months ago

@manutd_moon @Lingxiao234 @ZhennanJia13007 @HarryXu12 Guess what the robot has learned: What does it think?

Kun Lei11 months ago

Achieving a 100% success rate over 7 tasks. If you're troubled by insomnia, you can come and count for this demo.🤣 Timeline: Folding: 05:18-16:09 Juicing-placing: 16:09-19:09 Juicing-removal: 19:09-21:21 Pouring: 21:21-23:12 Unscrewing: 23:12-24:57 Bowling: 24:57-27:26 Push-T: 27:26-30:21 OverallJuicing: 30:21-31:57

Ted Xiao11 months ago

900/900 is a huge flex! Congrats on the strong results

Kun Lei11 months ago

Thanks, Ted!

lil’km11 months ago

Congrats Kun! Impressive results 🦾

Théo11 months ago

How is the lighting change robustness?

Kun Lei11 months ago

In real robot settings, we use a 3d point cloud as visual input without RGB. Thus, there will be no impact on performance from the changed lighting.

Alexander Soare11 months ago

This is amazing work. I love that the experiments section has an explicit focus on long-running uninterrupted success. Geeking out on the specifics here: I'm wondering about the choice of consistency distillation loss, particularly, running the teacher's full denoising chain rather than taking a single step as in the original CM paper.

Alan Zhuolun Zhao11 months ago

Very impressive results! Any reason for the different input/output action spaces? (Joint position vs delta EEF pose)

zkComRi.ETH11 months ago

RL-100 turns bots pro

Kun Lei11 months ago

Thanks!

pfung11 months ago

Great results, big congrats! 1. I was curious to see the fail cases in the few shot/one shot trials, are those in the embedded videos? (Didn't want to watch the whole vid to find 😄) 2. Any plans on releasing code for this? Thanks!!!

Kun Lei11 months ago

1. No problem, we will release the failure cases. 2. We will release it in the future, after the paper's acceptance.

pfung11 months ago

Thanks so much!!

pfung10 months ago

1 month ping :) wondering if you still plan to open source RL-100 code. In any case,congrats on a cool paper!

Antoni2 months ago

My take is that the whole trick is how to have a generic critic and how to reset the environment. Once you figure those two out, you can continuously improve on hundreds of tasks

Chi Chu11 months ago

100% success rate, that's unbelievable!🤯 Congrats Kun!

Kun Lei11 months ago

Thanks, Chi!

Crypto Kɑrhen11 months ago

insane uptime

Dominique Paul10 months ago

Why did you choose not to release the code for the paper?

sunny11 months ago

Great work! When would you publish the code？

MCS @ Safe Sentinel11 months ago

Congrats! Best thing I have seen this month. How does success rate compare on long horizon tasks vs shorter tasks? Hil-serl was bad at long horizon (>9a) but this seems to be way better

Related Videos

Introducing HiLMa-Res: a hierarchical RL framework for quadrupeds to tackle loco-manipulation tasks with sustained mobility! Designed for general learning tasks (vision-based, state-based, real-world data, etc), the robot now can step over stones🐾/navigate boxes📦/dribble⚽.

Introducing HiLMa-Res: a hierarchical RL framework for quadrupeds to tackle loco-manipulation tasks with sustained mobility! Designed for general learning tasks (vision-based, state-based, real-world data, etc), the robot now can step over stones🐾/navigate boxes📦/dribble⚽.

Zhongyu Li

11,720 views • 1 year ago

Excited to introduce SOLE-R1, a video-language reasoning model for zero-shot reward prediction for robot manipulation tasks! SOLE-R1 reasoning can serve as the SOLE signal for learning new tasks (completely from scratch) through online RL - i.e., robots start with random actions and learn previously unseen tasks guided only by SOLE-R1 rewards, without any demonstrations, ground-truth rewards, success indicators, or task-specific tuning. SOLE-R1 significantly outperforms strong baselines (e.g., Robometer, RoboReward, TOPReward, GPT-5, Gemini-3-Pro) in zero-shot online RL when evaluated across 40 tasks - including a real-world tabletop manipulation setting and 4 sim environments (LIBERO, ManiSkill, Meta-World, RoboSuite). We open source all models, training data, and code. Website, demos, and paper at: 🧵 (1/6)

Excited to introduce SOLE-R1, a video-language reasoning model for zero-shot reward prediction for robot manipulation tasks! SOLE-R1 reasoning can serve as the SOLE signal for learning new tasks (completely from scratch) through online RL - i.e., robots start with random actions and learn previously unseen tasks guided only by SOLE-R1 rewards, without any demonstrations, ground-truth rewards, success indicators, or task-specific tuning. SOLE-R1 significantly outperforms strong baselines (e.g., Robometer, RoboReward, TOPReward, GPT-5, Gemini-3-Pro) in zero-shot online RL when evaluated across 40 tasks - including a real-world tabletop manipulation setting and 4 sim environments (LIBERO, ManiSkill, Meta-World, RoboSuite). We open source all models, training data, and code. Website, demos, and paper at: 🧵 (1/6)

Philip Schroeder

19,275 views • 3 months ago

In order for robots to be deployed in the real world, performing tasks of real value, they must be reliable. Unfortunately, even more, most robotic demos work maybe 70-80% of the time at best. The way to get better reliability is to do real-world reinforcement learning: having the robot teach itself how to perform the task up to a high level of success. The key to doing this is to start with a core of expert human data, use that to train a policy then iteratively improve it, until finally finishing with on-policy reinforcement learning. Kun Lei talks through a unified framework for imitation and reinforcement learning based on PPO, which enables this improvement process. In this episode, Kun Lei explains the theory behind his reinforcement learning method and how it allowed his robot to run in a shopping mall juicing oranges for seven hours at a time, among experiments on a wide variety of tasks and embodiments. Watch episode 58 of RoboPapers now, hosted by Michael Cho - Rbt/Acc and Chris Paxton!

In order for robots to be deployed in the real world, performing tasks of real value, they must be reliable. Unfortunately, even more, most robotic demos work maybe 70-80% of the time at best. The way to get better reliability is to do real-world reinforcement learning: having the robot teach itself how to perform the task up to a high level of success. The key to doing this is to start with a core of expert human data, use that to train a policy then iteratively improve it, until finally finishing with on-policy reinforcement learning. Kun Lei talks through a unified framework for imitation and reinforcement learning based on PPO, which enables this improvement process. In this episode, Kun Lei explains the theory behind his reinforcement learning method and how it allowed his robot to run in a shopping mall juicing oranges for seven hours at a time, among experiments on a wide variety of tasks and embodiments. Watch episode 58 of RoboPapers now, hosted by Michael Cho - Rbt/Acc and Chris Paxton!

RoboPapers

18,813 views • 8 months ago

Robot policies must be both reliable and highly capable to be useful; the best way to achieve this level of performance is with reinforcement learning. However, for reinforcement learning you are usually stuck between two difficult options: reinforcement in the real world is often risky and expensive, while reinforcement learning in a traditional simulator takes a lot of engineering work and has a persistent sim-to-real gap. What if instead you could train your robot purely in a world model? RISE by Jiazhi Yang et al. uses a compositional world model to predict the future and evaluate progress. This allows for a self-improving pipeline, which learns a world model from real data and then learns how the robot should perform different tasks. This pipeline results in a data-driven way to improve policy performance from real data but without real-world reinforcement learning. Watch Episode #86 of RoboPapers, with Chris Paxton and Jiafei Duan, to learn more!

Robot policies must be both reliable and highly capable to be useful; the best way to achieve this level of performance is with reinforcement learning. However, for reinforcement learning you are usually stuck between two difficult options: reinforcement in the real world is often risky and expensive, while reinforcement learning in a traditional simulator takes a lot of engineering work and has a persistent sim-to-real gap. What if instead you could train your robot purely in a world model? RISE by Jiazhi Yang et al. uses a compositional world model to predict the future and evaluate progress. This allows for a self-improving pipeline, which learns a world model from real data and then learns how the robot should perform different tasks. This pipeline results in a data-driven way to improve policy performance from real data but without real-world reinforcement learning. Watch Episode #86 of RoboPapers, with Chris Paxton and Jiafei Duan, to learn more!

RoboPapers

38,772 views • 3 months ago

One more thing 🚀 Qwen’s agentic capability is no longer limited to the digital world — we’re bringing it into physical world. With our in-house robotic agentic system and navigation model, Qwen can now control a robot to execute tasks in real-world. #qwen #embodied #robotics

One more thing 🚀 Qwen’s agentic capability is no longer limited to the digital world — we’re bringing it into physical world. With our in-house robotic agentic system and navigation model, Qwen can now control a robot to execute tasks in real-world. #qwen #embodied #robotics

barry (xiong-hui) chen

25,112 views • 4 months ago

Egocentric Videos for Robot Training 🧵 Researchers at NYU and UC Berkeley published research where they developed a system called EgoZero that trains robots using human demonstration videos recorded with smart glasses. 👓 The system converts first-person human actions into 3D point-based state-action representations. The policies executed on a gripper-equipped robot, achieved a 70% zero-shot success rate across seven manipulation tasks, with only 20 minutes of human data per task. EgoZero stands as one of the solid empirical proofs that egocentric smart-glass video collected from everyday human behavior can serve as powerful, scalable training data for real robot learning. Vincent Liu Ademi Adeniji

Egocentric Videos for Robot Training 🧵 Researchers at NYU and UC Berkeley published research where they developed a system called EgoZero that trains robots using human demonstration videos recorded with smart glasses. 👓 The system converts first-person human actions into 3D point-based state-action representations. The policies executed on a gripper-equipped robot, achieved a 70% zero-shot success rate across seven manipulation tasks, with only 20 minutes of human data per task. EgoZero stands as one of the solid empirical proofs that egocentric smart-glass video collected from everyday human behavior can serve as powerful, scalable training data for real robot learning. Vincent Liu Ademi Adeniji

VaderResearch

23,055 views • 1 year ago

A big part of scaling robot learning to solve real-world problems is that we somehow need to get enough diverse, high-quality data to train our robots to perform useful things. GPT and its fellow large language models were bootstrapped and proved out on a massive dataset of real-world language data. Unfortunately, despite our best efforts, similarly massive datasets don’t really exist for robotics — so, in our unending pursuit of high-quality, useful data, we turn to simulation. I compared a couple recent works on sim-to-real robot manipulation, which discuss how to train perception-driven manipulation policies in simulation, in such a way that they’re useful in the real world. - DextraH-RGB, from NVIDIA - Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation, also from NVIDIA — specifically the GEAR lab - Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids, another GEAR lab paper - Local Policies Enable Zero-shot Long-Horizon Manipulation, from CMU (video from DextrAH-RGB)

A big part of scaling robot learning to solve real-world problems is that we somehow need to get enough diverse, high-quality data to train our robots to perform useful things. GPT and its fellow large language models were bootstrapped and proved out on a massive dataset of real-world language data. Unfortunately, despite our best efforts, similarly massive datasets don’t really exist for robotics — so, in our unending pursuit of high-quality, useful data, we turn to simulation. I compared a couple recent works on sim-to-real robot manipulation, which discuss how to train perception-driven manipulation policies in simulation, in such a way that they’re useful in the real world. - DextraH-RGB, from NVIDIA - Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation, also from NVIDIA — specifically the GEAR lab - Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids, another GEAR lab paper - Local Policies Enable Zero-shot Long-Horizon Manipulation, from CMU (video from DextrAH-RGB)

Chris Paxton

20,486 views • 1 year ago

What if robots could improve themselves by learning from their own failures in the real-world? Introducing 𝗣𝗟𝗗 (𝗣𝗿𝗼𝗯𝗲, 𝗟𝗲𝗮𝗿𝗻, 𝗗𝗶𝘀𝘁𝗶𝗹𝗹) — a recipe that enables Vision-Language-Action (VLA) models to self-improve for high-precision manipulation tasks. PLD couples real-world residual reinforcement learning with standard supervised fine-tuning — letting robots discover, recover, and distill their own data flywheel. Quick 🧵

What if robots could improve themselves by learning from their own failures in the real-world? Introducing 𝗣𝗟𝗗 (𝗣𝗿𝗼𝗯𝗲, 𝗟𝗲𝗮𝗿𝗻, 𝗗𝗶𝘀𝘁𝗶𝗹𝗹) — a recipe that enables Vision-Language-Action (VLA) models to self-improve for high-precision manipulation tasks. PLD couples real-world residual reinforcement learning with standard supervised fine-tuning — letting robots discover, recover, and distill their own data flywheel. Quick 🧵

Wenli Xiao

185,865 views • 10 months ago

Deployment is only the first step. You need to learn from it and take the next steps toward positive ROI. Our journey: Step 1. Nine month ago, our robot already assembled a complex real product for our customer, from start to finish. But it wasn’t reliable. Step 2. Our robot performed a complex benchmark task reliably 100 times in a row. Step 3. (Completing) Our robot assembling a complex real product from start to finish 100+ times in a row.

Deployment is only the first step. You need to learn from it and take the next steps toward positive ROI. Our journey: Step 1. Nine month ago, our robot already assembled a complex real product for our customer, from start to finish. But it wasn’t reliable. Step 2. Our robot performed a complex benchmark task reliably 100 times in a row. Step 3. (Completing) Our robot assembling a complex real product from start to finish 100+ times in a row.

Igor Kulakov (MicroFactory)

11,428 views • 1 month ago

Haven't been to a conference in a while, really excited to be at #NeurIPS2024! I'll be helping present 4 of our group's recent papers: 1. Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL 2. Distributional Successor Features Enable Zero-Shot Policy Optimization 3. Learning to Cooperate with Humans using Generative Agents 4. Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning Find more details on each paper and where to find us in this thread (1/6)

Haven't been to a conference in a while, really excited to be at #NeurIPS2024! I'll be helping present 4 of our group's recent papers: 1. Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL 2. Distributional Successor Features Enable Zero-Shot Policy Optimization 3. Learning to Cooperate with Humans using Generative Agents 4. Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning Find more details on each paper and where to find us in this thread (1/6)

Abhishek Gupta

10,803 views • 1 year ago

We are open-sourcing Wall-OSS-0.5. Pretrain Once, Act Anywhere. Wall-OSS-0.5 is a VLA model for real-world robotic manipulation, exploring whether pretraining alone can produce robot capabilities directly testable on physical hardware before task-specific fine-tuning. Key technical highlights: • Gradient-bridged co-training • Vision-Aligned RVQ Action Tokenizer • Action-Space Supervision • DMuon distributed optimizer In zero-shot real-robot evaluation, the pretrained checkpoint achieved task-progress scores above 80 on multiple tasks, including Block Sorting, Fruit Sorting, Ring Stacking, and Rope Tightening. Paper, code, blog, and uncut videos:

We are open-sourcing Wall-OSS-0.5. Pretrain Once, Act Anywhere. Wall-OSS-0.5 is a VLA model for real-world robotic manipulation, exploring whether pretraining alone can produce robot capabilities directly testable on physical hardware before task-specific fine-tuning. Key technical highlights: • Gradient-bridged co-training • Vision-Aligned RVQ Action Tokenizer • Action-Space Supervision • DMuon distributed optimizer In zero-shot real-robot evaluation, the pretrained checkpoint achieved task-progress scores above 80 on multiple tasks, including Block Sorting, Fruit Sorting, Ring Stacking, and Rope Tightening. Paper, code, blog, and uncut videos:

X Square Robot

24,659 views • 3 months ago

$We’re excited to launch Instance – a success detector for robot rollouts! Describe your task and drop in a dataset, and it judges success/fail on each rollout with detailed subtask captions– it’s more accurate than Claude Opus 4.8, at a fraction of the latency. This is just the first step toward our larger goal: an autonomous evaluation rig for robot learning. Today, evaluating a robot policy means humans have to watch the robot roll out, mark success, and reset the scene– we're automating that, starting with the success judge.$

We’re excited to launch Instance – a success detector for robot rollouts! Describe your task and drop in a dataset, and it judges success/fail on each rollout with detailed subtask captions– it’s more accurate than Claude Opus 4.8, at a fraction of the latency. This is just the first step toward our larger goal: an autonomous evaluation rig for robot learning. Today, evaluating a robot policy means humans have to watch the robot roll out, mark success, and reset the scene– we're automating that, starting with the success judge.

Lucy Cai

73,929 views • 2 months ago

🚨🇺🇸FIGURE ROBOT LEARNS TO WALK IN HOURS Turns out you can teach a robot to walk—if you’ve got a few years to burn in a physics simulator (or a few hours of GPU time). Figure’s team trained a neural network using reinforcement learning to mimic natural human locomotion, and the wild part is it worked on the first try with real hardware. No tuning. No fiddling. Just plug and strut. Thanks to sim-to-real transfer and some sneaky domain randomization, the robot didn’t just wobble out of the lab—it walked like it had somewhere to be. Next up: running. Not terrifying at all. Source: Figure

🚨🇺🇸FIGURE ROBOT LEARNS TO WALK IN HOURS Turns out you can teach a robot to walk—if you’ve got a few years to burn in a physics simulator (or a few hours of GPU time). Figure’s team trained a neural network using reinforcement learning to mimic natural human locomotion, and the wild part is it worked on the first try with real hardware. No tuning. No fiddling. Just plug and strut. Thanks to sim-to-real transfer and some sneaky domain randomization, the robot didn’t just wobble out of the lab—it walked like it had somewhere to be. Next up: running. Not terrifying at all. Source: Figure

Mario Nawfal

117,866 views • 1 year ago

🧐Applying world models to improve real-world policy on challenging manipulation tasks used to be considered out of reach. 😌After sustained effort, we’re now seeing encouraging progress. 🚀Thrilled to introduce RISE: Self-Improving Robot Policy with Compositional World Model RISE is, to our knowledge, the first work to use a world model as an effective learning environment for challenging real-world manipulation, enabling policy improvement on tasks that demand high dynamics, dexterity, and precision. Incredible teamwork with KL Hongyang Li Yue Xiangyu Hao Zhao Chonghao Sima

🧐Applying world models to improve real-world policy on challenging manipulation tasks used to be considered out of reach. 😌After sustained effort, we’re now seeing encouraging progress. 🚀Thrilled to introduce RISE: Self-Improving Robot Policy with Compositional World Model RISE is, to our knowledge, the first work to use a world model as an effective learning environment for challenging real-world manipulation, enabling policy improvement on tasks that demand high dynamics, dexterity, and precision. Incredible teamwork with KL Hongyang Li Yue Xiangyu Hao Zhao Chonghao Sima

Jiazhi Yang

77,124 views • 7 months ago

So I heard we need more data for robot learning :) Purely real world teleop is expensive and slow, making large scale data collection challenging. I’ve been excited about getting more data into robot learning, going beyond just real-world teleop data. To this end, we’ve been scaling up data generation with RL in realistic simulations generated on the fly from crowdsourced videos. Enables realistic data collection, much more cheaply than purely real world teleop. Importantly, data collection becomes even*cheaper* with more environments, allowing training with over 100x more data. Transfers to real robots for generalizable manipulation. A 🧵 (1/N)

So I heard we need more data for robot learning :) Purely real world teleop is expensive and slow, making large scale data collection challenging. I’ve been excited about getting more data into robot learning, going beyond just real-world teleop data. To this end, we’ve been scaling up data generation with RL in realistic simulations generated on the fly from crowdsourced videos. Enables realistic data collection, much more cheaply than purely real world teleop. Importantly, data collection becomes evencheaper with more environments, allowing training with over 100x more data. Transfers to real robots for generalizable manipulation. A 🧵 (1/N)

Abhishek Gupta

13,358 views • 1 year ago