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.

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

1,132 subscribers

88,312 views • 3 years ago •via X (Twitter)

Science & Technology Education

Anya Rossi• Live Now

Private livecam show

6 Comments

Jingyun Yang3 years ago

The pre-train + fine-tune paradigm in ML has shown promise in a lot of domains. RoboFuME is a reset-free fine-tuning system that trains with data from an existing robot dataset collected at a different institution and takes just 3 hours to fine-tune autonomously. [1/6]

Jingyun Yang3 years ago

In pre-training, we use diverse prior demonstration data from the Bridge V2 dataset, which was collected at a different institution. For each task, we pick about 1,000 relevant trajectories in the dataset as prior data for the agent. [2/6]

Jingyun Yang3 years ago

In addition to the prior data, we spend 30 minutes to collect 50 successful demos for both the forward and backward target tasks, as well as 10 minutes of failure demos for both tasks. [3/6]

Jingyun Yang3 years ago

We pre-train an RL policy from both the prior data and the target demonstration data using an offline RL algorithm. Then, we rollout the policy to perform the forward and backward tasks alternatively while updating the policy with a VLM reward model as reward feedback. [4/6]

Jingyun Yang3 years ago

We test our system on 5 real robot tasks. Our method significantly improves over both offline-only and BC performance after 30k steps of online interaction (2-4 hours). [5/6]

Jingyun Yang3 years ago

This is joint work with an amazing team including @maxsobolmark, @brandontvu, @archit_sharma97, @leto__jean, @chelseabfinn Website: Paper: [6/6]

Related Videos

A reward model that works, zero-shot, across robots, tasks, and scenes? Introducing Robometer: Scaling general-purpose robotic reward models with 1M+ trajectories. Enables zero-shot: online/offline/model-based RL, data retrieval + IL, automatic failure detection, and more! 🧵 (1/12)

A reward model that works, zero-shot, across robots, tasks, and scenes? Introducing Robometer: Scaling general-purpose robotic reward models with 1M+ trajectories. Enables zero-shot: online/offline/model-based RL, data retrieval + IL, automatic failure detection, and more! 🧵 (1/12)

Jesse Zhang ✈️ CoRL 2026

103,597 views • 8 months ago

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

LeRobot

22,592 views • 3 months ago

Let’s do 🍒 Cherry Picking with Reinforcement Learning - 🥢 Dynamic fine manipulation with chopsticks - 🤖 Only 30 minutes of real world interactions - ⛔️ Too lazy for parameter tuning = off-the-shelf RL algo + default params + 3 seeds in real world

Let’s do 🍒 Cherry Picking with Reinforcement Learning - 🥢 Dynamic fine manipulation with chopsticks - 🤖 Only 30 minutes of real world interactions - ⛔️ Too lazy for parameter tuning = off-the-shelf RL algo + default params + 3 seeds in real world

Kay - Liyiming Ke @ CoRL2026

51,436 views • 3 years ago

Excited to share our #ICML2023 paper ✨LIV✨! Extending VIP, LIV is at once a pre-training, fine-tuning, and (zero-shot!) multi-modal reward method for (real-world!) language-conditioned robotic control. Project: Code & Model: 🧵:

Excited to share our #ICML2023 paper ✨LIV✨! Extending VIP, LIV is at once a pre-training, fine-tuning, and (zero-shot!) multi-modal reward method for (real-world!) language-conditioned robotic control. Project: Code & Model: 🧵:

Jason Ma

55,068 views • 3 years ago

🤖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 views • 6 months ago

Instead of asking a VLM to output progress, it reads the model’s internal belief directly from token logits. No in-context learning. No fine-tuning. No reward training. 📈 We introduce: TOPReward, a zero-shot reward modeling approach for robotics using token probabilities from pretrained video VLMs. The simplest way of doing reward modelling for robotics! Project: 🧵👇

Instead of asking a VLM to output progress, it reads the model’s internal belief directly from token logits. No in-context learning. No fine-tuning. No reward training. 📈 We introduce: TOPReward, a zero-shot reward modeling approach for robotics using token probabilities from pretrained video VLMs. The simplest way of doing reward modelling for robotics! Project: 🧵👇

Jiafei Duan

108,387 views • 8 months ago

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

13,659 views • 1 year ago

Meet BFM-Zero: A Promptable Humanoid Behavioral Foundation Model w/ Unsupervised RL👉 🧩ONE latent space for ALL tasks ⚡Zero-shot goal reaching, tracking, and reward optimization (any reward at test time), from ONE policy 🤖Natural recovery & transition

Meet BFM-Zero: A Promptable Humanoid Behavioral Foundation Model w/ Unsupervised RL👉 🧩ONE latent space for ALL tasks ⚡Zero-shot goal reaching, tracking, and reward optimization (any reward at test time), from ONE policy 🤖Natural recovery & transition

Yitang Li

84,263 views • 1 year ago

Can robots self-improve by collecting data autonomously🤖? Introducing SOAR: a system for large-scale autonomous data collection 🚀 and autonomous improvement📈of a multi-task language-conditioned policy in diverse scenes without human interventions .

Can robots self-improve by collecting data autonomously🤖? Introducing SOAR: a system for large-scale autonomous data collection 🚀 and autonomous improvement📈of a multi-task language-conditioned policy in diverse scenes without human interventions .

Paul Zhou

47,739 views • 2 years ago

We asked Sholto Douglas from Anthropic about the costs of RL (Reinforcement Learning) runs. "In Dario Amodei's essay, he said that RL runs cost only $1M back in December." "RL is a more naively parallelizable and scalable than pre-training." "With pre-training, you need everything in one big data center ideally. For RL, in theory, you could scale all over the world."

We asked Sholto Douglas from Anthropic about the costs of RL (Reinforcement Learning) runs. "In Dario Amodei's essay, he said that RL runs cost only $1M back in December." "RL is a more naively parallelizable and scalable than pre-training." "With pre-training, you need everything in one big data center ideally. For RL, in theory, you could scale all over the world."

TBPN

76,696 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 ✈️ CoRL 2026

19,852 views • 5 months ago

Our model can now learn from its own experience with RL! Our new π*0.6 model can more than double throughput over a base model trained without RL, and can perform real-world tasks: making espresso drinks, folding diverse laundry, and assembling boxes. More in the thread below.

Our model can now learn from its own experience with RL! Our new π*0.6 model can more than double throughput over a base model trained without RL, and can perform real-world tasks: making espresso drinks, folding diverse laundry, and assembling boxes. More in the thread below.

Physical Intelligence

713,421 views • 11 months ago

💡Divergence thinking💡 is a hallmark of human creativity and problem-solving 🤖Can LLMs also do divergent reasoning to generate diverse solutions🤔? Introducing Flow-of-Reasoning (FoR) 🌊, a data-efficient way of training LLM policy to generate diverse, high-quality reasoning trajectories Unlike existing RL (like PPO) and planning (like MCTS) to find the max-reward trajectory (akin to convergent thinking), FoR connects LLM reasoning with the #GFlowNet formulation and enables LLMs to find trajectories proportional to reward distribution. 🎬The demo video illustrates how FoR learns and infers multiple solutions to a ♠️Game24 puzzle. 🎯Inferring for diverse solutions could be useful for robustness, data augmentation, and enhanced model generalization. Project page: Paper: Github:

💡Divergence thinking💡 is a hallmark of human creativity and problem-solving 🤖Can LLMs also do divergent reasoning to generate diverse solutions🤔? Introducing Flow-of-Reasoning (FoR) 🌊, a data-efficient way of training LLM policy to generate diverse, high-quality reasoning trajectories Unlike existing RL (like PPO) and planning (like MCTS) to find the max-reward trajectory (akin to convergent thinking), FoR connects LLM reasoning with the #GFlowNet formulation and enables LLMs to find trajectories proportional to reward distribution. 🎬The demo video illustrates how FoR learns and infers multiple solutions to a ♠️Game24 puzzle. 🎯Inferring for diverse solutions could be useful for robustness, data augmentation, and enhanced model generalization. Project page: Paper: Github:

Lianhui Qin

50,493 views • 2 years ago

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

454,623 views • 7 months ago

The best way to get robust, high-quality robot performance is through reinforcement learning; but RL in either the real world or a traditional simulation has lots of limitations. Instead, Jiazhi Yang in RISE does RL in a compositional world model. Learn more ->

The best way to get robust, high-quality robot performance is through reinforcement learning; but RL in either the real world or a traditional simulation has lots of limitations. Instead, Jiazhi Yang in RISE does RL in a compositional world model. Learn more ->

Chris Paxton

34,102 views • 5 months ago

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 collaboration with Siemens, led by Brian Zhu, Momen Khalil, Emanuele Poggi from Siemens and E Harrison from Berkeley, with lots of amazing contributors!

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 collaboration with Siemens, led by Brian Zhu, Momen Khalil, Emanuele Poggi from Siemens and E Harrison from Berkeley, with lots of amazing contributors!

Sergey Levine

47,972 views • 1 month ago

(1/5) Everyone says dexterous manipulation needs real-world data. These policies were trained with none. Our robots mesh gears, thread nuts, and insert pegs zero-shot from RGB, and the same recipe works for locomotion across terrains much harder than before. The missing piece wasn't a better RL algorithm, but a small tweak to what the robot practices. Today we're releasing Success-Guided Sampling (SGS), a simple change to the task sampler that unlocks new levels of dexterity and agility with large-scale, end-to-end sim2real RL. No demonstrations, no per-task reward tuning. Just PPO at scale. Check out the policies yourself at All videos at 1x speed! 🧵 🔈

(1/5) Everyone says dexterous manipulation needs real-world data. These policies were trained with none. Our robots mesh gears, thread nuts, and insert pegs zero-shot from RGB, and the same recipe works for locomotion across terrains much harder than before. The missing piece wasn't a better RL algorithm, but a small tweak to what the robot practices. Today we're releasing Success-Guided Sampling (SGS), a simple change to the task sampler that unlocks new levels of dexterity and agility with large-scale, end-to-end sim2real RL. No demonstrations, no per-task reward tuning. Just PPO at scale. Check out the policies yourself at All videos at 1x speed! 🧵 🔈

Octi Zhang

212,309 views • 1 month ago

Real-world robot data is expensive and slow to collect, creating a major challenge for humanoid development. 🤖 The NVIDIA GR00T N1.6 open vision language action model is pre-trained on a diverse mix of data, including thousands of hours of Stanford Vision and Learning Lab’s BEHAVIOR simulation data, which covers long-horizon everyday manipulation tasks. This diverse training is the key to robust cross-embodiment performance and real-world adaptability. 🌍 Read the blog 🔗

Real-world robot data is expensive and slow to collect, creating a major challenge for humanoid development. 🤖 The NVIDIA GR00T N1.6 open vision language action model is pre-trained on a diverse mix of data, including thousands of hours of Stanford Vision and Learning Lab’s BEHAVIOR simulation data, which covers long-horizon everyday manipulation tasks. This diverse training is the key to robust cross-embodiment performance and real-world adaptability. 🌍 Read the blog 🔗

NVIDIA Robotics

13,562 views • 9 months ago

Announcing ManiSkill2: a unified, fast, and accessible benchmark - robot learning made simple! - ✅Pip installable & easily deployable - ⚡Blazingly fast visual RL support - ✨Diverse task families, objects, & 4M demonstration frames - 🖥️Interactive GUI

Announcing ManiSkill2: a unified, fast, and accessible benchmark - robot learning made simple! - ✅Pip installable & easily deployable - ⚡Blazingly fast visual RL support - ✨Diverse task families, objects, & 4M demonstration frames - 🖥️Interactive GUI

Hao Su Lab

42,123 views • 3 years ago