Sergey Levine's banner
Sergey Levine's profile picture

Sergey Levine

@svlevine • 139,882 subscribers

Associate Professor at UC Berkeley Co-founder, Physical Intelligence

Shorts

Tmrw(Sat) I'll talk about pi-zero & HIL-SERL at the CoRL workshop on Manipulation in a World of Abundant data at 9 am CET (Jupiter room)! Also leading discussion about cross-embodiment in the X-Embodiment workshop at 9:30 am (Terra) Come find out about cool recent advances!

Tmrw(Sat) I'll talk about pi-zero & HIL-SERL at the CoRL workshop on Manipulation in a World of Abundant data at 9 am CET (Jupiter room)! Also leading discussion about cross-embodiment in the X-Embodiment workshop at 9:30 am (Terra) Come find out about cool recent advances!

43,171 views

Videos

LIVE
1.2k

Anya Rossi

sweetdream.ai

SweetDream.aiSponsored•Livecam
Streaming Now

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

HD live stream
Exclusive private shows
1.2k viewers online

Free preview available • Premium content

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!
3:21
svlevine's profile picture

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

46,467 views • 1 month ago

Flow reversal steering allows "steering" diffusion-based VLAs with high-level actions, for example from VLM reasoning. This also lets us run RL in the diffusion noise space with exploration guided by high-level reasoning: think through a task, then practice it! 👇
0:55
svlevine's profile picture

Flow reversal steering allows "steering" diffusion-based VLAs with high-level actions, for example from VLM reasoning. This also lets us run RL in the diffusion noise space with exploration guided by high-level reasoning: think through a task, then practice it! 👇

Sergey Levine

90,281 views • 4 months ago

If you have a policy that uses diffusion/flow (e.g. diffusion VLA), you can run RL where the actor chooses the noise, which is then denoised by the policy to produce an action. This method, which we call diffusion steering (DSRL), leads to a remarkably efficient RL method! 🧵👇
0:39
svlevine's profile picture

If you have a policy that uses diffusion/flow (e.g. diffusion VLA), you can run RL where the actor chooses the noise, which is then denoised by the policy to produce an action. This method, which we call diffusion steering (DSRL), leads to a remarkably efficient RL method! 🧵👇

Sergey Levine

153,232 views • 1 year ago

We finished evaluating π0.7, our new model at Physical Intelligence. What I'm most excited about with π0.7 is that it's starting to show some surprising emergent compositional generalization, being able to both perform complex tasks and learn new tasks just from instructions.
1:07
svlevine's profile picture

We finished evaluating π0.7, our new model at Physical Intelligence. What I'm most excited about with π0.7 is that it's starting to show some surprising emergent compositional generalization, being able to both perform complex tasks and learn new tasks just from instructions.

Sergey Levine

61,590 views • 6 months ago

A while back Benjie Holson described a set of "Robot Olympics" challenge tasks -- washing a pan, making a peanut butter sandwich, and more. We tried to fine-tune our models at PI to these tasks, and found that we could do most of them. A few highlights below.
1:12
svlevine's profile picture

A while back Benjie Holson described a set of "Robot Olympics" challenge tasks -- washing a pan, making a peanut butter sandwich, and more. We tried to fine-tune our models at PI to these tasks, and found that we could do most of them. A few highlights below.

Sergey Levine

81,429 views • 10 months ago

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
0:46
svlevine's profile picture

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

38,514 views • 4 months ago

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!
1:22
svlevine's profile picture

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

32,162 views • 3 months ago

Really excited to share what I've been working on with my colleagues at Physical Intelligence! We've developed a prototype robotic foundation model that can fold laundry, assemble a box, bus a table, and many other things. We've written a paper and blog post about it. 🧵👇
5:10
svlevine's profile picture

Really excited to share what I've been working on with my colleagues at Physical Intelligence! We've developed a prototype robotic foundation model that can fold laundry, assemble a box, bus a table, and many other things. We've written a paper and blog post about it. 🧵👇

Sergey Levine

115,143 views • 2 years ago

One of my favorite results for π*0.6 is this video: 13 hours of making lattes, Americanos, and espresso for folks at our office in San Francisco.
9:42
svlevine's profile picture

One of my favorite results for π*0.6 is this video: 13 hours of making lattes, Americanos, and espresso for folks at our office in San Francisco.

Sergey Levine

57,905 views • 11 months ago

Real-time inference is a big challenge for VLAs. We’ve been working on a way to amortize inference delays in π0.5. Our new Real-Time Chunking (RTC) method speeds up π0.5 by allowing the robot to “think” while it’s moving, which makes it quite a bit faster! 🧵👇
0:29
svlevine's profile picture

Real-time inference is a big challenge for VLAs. We’ve been working on a way to amortize inference delays in π0.5. Our new Real-Time Chunking (RTC) method speeds up π0.5 by allowing the robot to “think” while it’s moving, which makes it quite a bit faster! 🧵👇

Sergey Levine

73,796 views • 1 year ago

Diffusion models make great images. But can they drive robots? Usually that gets complicated really fast. We figured out how to get a Stable Diffusion model (based on Instruct pix2pix) to drive robotic instruction following. Simple recipe, works on a wide range of tasks. Thread👇
0:29
svlevine's profile picture

Diffusion models make great images. But can they drive robots? Usually that gets complicated really fast. We figured out how to get a Stable Diffusion model (based on Instruct pix2pix) to drive robotic instruction following. Simple recipe, works on a wide range of tasks. Thread👇

Sergey Levine

126,572 views • 3 years ago

Very happy to announce that we are open-sourcing the π₀ model, weights, and some fine-tuned checkpoints! Hoping this leads to lots of great follow-up research: Here is a fun test from our friends at UPenn.
0:31
svlevine's profile picture

Very happy to announce that we are open-sourcing the π₀ model, weights, and some fine-tuned checkpoints! Hoping this leads to lots of great follow-up research: Here is a fun test from our friends at UPenn.

Sergey Levine

72,593 views • 1 year ago

Watch this robot dog learn to walk from scratch in real time! Our new method, APRL, dynamically adjusts exploration constraints to enable fast and performant RL directly in the real world. APRL can also adapt to changes in the terrain. No simulation, no demos. A thread 👇
0:57
svlevine's profile picture

Watch this robot dog learn to walk from scratch in real time! Our new method, APRL, dynamically adjusts exploration constraints to enable fast and performant RL directly in the real world. APRL can also adapt to changes in the terrain. No simulation, no demos. A thread 👇

Sergey Levine

105,579 views • 3 years ago

At Physical Intelligence, we teamed up with Weave Robotics and Ultra to stress-test our models in real-world deployments. This was a really fun collaboration that saw our latest pi06 model running in production at Sea Breeze Cleaners and a real warehouse! More below.
19:06
svlevine's profile picture

At Physical Intelligence, we teamed up with Weave Robotics and Ultra to stress-test our models in real-world deployments. This was a really fun collaboration that saw our latest pi06 model running in production at Sea Breeze Cleaners and a real warehouse! More below.

Sergey Levine

32,352 views • 8 months ago

We trained a robotic foundation model that can drive mobile robots in six different countries, and navigate Sproul Plaza in midday on the UC Berkeley campus! Some cool new work w/ noriaki_hirose, Lydia Ignatova, Kyle Stachowicz, Catherine Glossop, Dhruv Shah
1:55
svlevine's profile picture

We trained a robotic foundation model that can drive mobile robots in six different countries, and navigate Sproul Plaza in midday on the UC Berkeley campus! Some cool new work w/ noriaki_hirose, Lydia Ignatova, Kyle Stachowicz, Catherine Glossop, Dhruv Shah

Sergey Levine

52,979 views • 1 year ago

Self-supervised representation learning looks a bit like RL. What if we literally use RL as a SSL method for visual representations? Turns out that it works quite well. In new work by Dibya Ghosh, we show how this can be done:
1:09
svlevine's profile picture

Self-supervised representation learning looks a bit like RL. What if we literally use RL as a SSL method for visual representations? Turns out that it works quite well. In new work by Dibya Ghosh, we show how this can be done:

Sergey Levine

48,840 views • 1 year ago

Language following is a tough problem for VLAs: while these models can follow complex language, in practice getting datasets that enable language following is hard. We developed a method to counterfactually and automatically label data to improve language following! 🧵👇
1:12
svlevine's profile picture

Language following is a tough problem for VLAs: while these models can follow complex language, in practice getting datasets that enable language following is hard. We developed a method to counterfactually and automatically label data to improve language following! 🧵👇

Sergey Levine

44,287 views • 1 year ago

Benjie posted a reply to the Physical Intelligence "Olympics" attempt: Some nice discussion about what makes a task hard, how current learning methods should change how people think about robotic capability, and some interesting commentary.
0:33
svlevine's profile picture

Benjie posted a reply to the Physical Intelligence "Olympics" attempt: Some nice discussion about what makes a task hard, how current learning methods should change how people think about robotic capability, and some interesting commentary.

Sergey Levine

27,380 views • 9 months ago

Scaling laws in deep RL? Turns out that batch size, learning rate, and UTD (update-to-data) for getting the most efficient and scalable deep RL has predictable relationships. Checkout the analysis in new work by Oleg Rybkin & collaborators:
0:29
svlevine's profile picture

Scaling laws in deep RL? Turns out that batch size, learning rate, and UTD (update-to-data) for getting the most efficient and scalable deep RL has predictable relationships. Checkout the analysis in new work by Oleg Rybkin & collaborators:

Sergey Levine

43,494 views • 1 year ago