Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

Robot AI brains, aka Vision-Language-Action models, cannot adapt to new tasks as easily as LLMs like Gemini, ChatGPT, or Grok. LLMs can adapt quickly with their in-context learning (ICL) capabilities. But can we inject ICL abilities into a pre-trained VLA like pi0? Yes! Introducing RICL (Retraining for In-Context Learning),... show more

Kaustubh Sridhar

1,641 subscribers

52,897 Aufrufe • vor 1 Jahr •via X (Twitter)

Wissenschaft & Technologie Bildung

Anya Rossi• Live Now

Private livecam show

9 Kommentare

Profilbild von Kaustubh Sridhar

Kaustubh Sridharvor 1 Jahr

How does it work? RICL-pi0 is given access to a few demos in new/long-tail tasks. It never finetunes on these demos… only retrieves relevant bits to throw into its context in each timestep. We trained it on a few “priming” tasks, with RAG during training, so that it learns to use its retrieved context effectively. We then evaluate on unseen/long-tail tasks. A video with comparisons of pi0 and RICL-pi0:

Profilbild von Kaustubh Sridhar

Kaustubh Sridharvor 1 Jahr

“Finetuning like you pretrain”: If you finetune on target task data exactly like RICL was pretrained on the priming tasks, you get the best performing policies. The below video shows the reactivity and robustness of RICL-pi0-finetuned policies in various target tasks, under human perturbations

Profilbild von Kaustubh Sridhar

Kaustubh Sridharvor 1 Jahr

TLDR: RICL-pi0-finetuned >> pi0-finetuned ≈ RICL-pi0 (only RAG + ICL) >> pi0 (out-of-the-box) Please find more information about RICL (including many limitations) at website: and arxiv: This work was my last PhD paper before joining @GoogleDeepmind and would not have been possible without @1_souradeep, @dineshjayaraman, and @insuplee. RICL has been accepted to CoRL 2025 in Seoul, Korea. If you’re there, @1_souradeep will be happy to tell you all about our work.

Profilbild von Kaustubh Sridhar

Kaustubh Sridharvor 1 Jahr

FYI, this work builds on our previous work REGENT (qt’ed below):

Profilbild von PrismaX

PrismaXvor 1 Jahr

Pi0 unlocked the promptability upgrade! Let's give it to 1,000 tele-op users and see how robots learn with real-world feedback.

Profilbild von Mykhailo Sorochuk

Mykhailo Sorochukvor 1 Jahr

Impressive approach. RAG + ICL for VLA? That could really push robot adaptability. Curious how well it generalizes without fine-tuning.

Profilbild von Aadarsh R

Aadarsh Rvor 1 Jahr

ICL can unlock test-time diverse generalization. Very interesting direction! Nice to see Idli plate in the experiments :)

Profilbild von Abhranil Chandra @ RLC 2026 🇨🇦

Abhranil Chandra @ RLC 2026 🇨🇦vor 1 Jahr

Great work @_k_sridhar. Both this and REGENT, seem like an interesting direction for fast adaptation in FMs for embodied tasks...

Profilbild von KillerBambi2

KillerBambi2vor 1 Jahr

AI brain is not the same as a human brain

Ähnliche Videos

Vision-language models perform diverse tasks via in-context learning. Time for robots to do the same! Introducing In-Context Robot Transformer (ICRT): a robot policy that learns new tasks by prompting with robot trajectories, without any fine-tuning. [1/N]

Vision-language models perform diverse tasks via in-context learning. Time for robots to do the same! Introducing In-Context Robot Transformer (ICRT): a robot policy that learns new tasks by prompting with robot trajectories, without any fine-tuning. [1/N]

Max Fu

40,466 Aufrufe • vor 2 Jahren

Apart from solving new tasks, memory also allows our policies to be more robust: we show early signs of in-context adaptation, where the robot learns to adapt its behavior on-the-fly by learning from its past mistakes.

Apart from solving new tasks, memory also allows our policies to be more robust: we show early signs of in-context adaptation, where the robot learns to adapt its behavior on-the-fly by learning from its past mistakes.

Physical Intelligence

14,923 Aufrufe • vor 7 Monaten

In-context learning is the holy grail of robot learning. It is challenging because: 1. long-context training and infra (ICL can easily max out disk I/O) 2. need model to follow multimodal (sensorimotor) condition 3. data collection strategy and how to pair data Tried to get it work in 2024, where we proposed in-context robot transformer, an autoregressive formulation: Congrats on the launch! Looking forward to seeing few-shot improvement via in-context learning in the future!

In-context learning is the holy grail of robot learning. It is challenging because: 1. long-context training and infra (ICL can easily max out disk I/O) 2. need model to follow multimodal (sensorimotor) condition 3. data collection strategy and how to pair data Tried to get it work in 2024, where we proposed in-context robot transformer, an autoregressive formulation: Congrats on the launch! Looking forward to seeing few-shot improvement via in-context learning in the future!

Max Fu

10,664 Aufrufe • vor 2 Monaten

Excited to finally share Generative Value Learning (GVL), my Google DeepMind project on extracting universal value functions from long-context VLMs via in-context learning! We discovered a simple method to generate zero-shot and few-shot values for 300+ robot tasks and 50+ datasets using SOTA VLMs like Gemini (Try out the demo on our website on your robot video today!) I worked a lot on leveraging foundation models as guidance for robots in my PhD, and to me, this result forges a new frontier in how we can use foundation models for robot learning, given its broad applicability independent of embodiment and task types. Quite excited about how we can build on this work as a community!

Excited to finally share Generative Value Learning (GVL), my Google DeepMind project on extracting universal value functions from long-context VLMs via in-context learning! We discovered a simple method to generate zero-shot and few-shot values for 300+ robot tasks and 50+ datasets using SOTA VLMs like Gemini (Try out the demo on our website on your robot video today!) I worked a lot on leveraging foundation models as guidance for robots in my PhD, and to me, this result forges a new frontier in how we can use foundation models for robot learning, given its broad applicability independent of embodiment and task types. Quite excited about how we can build on this work as a community!

Jason Ma

98,270 Aufrufe • vor 1 Jahr

✨ Introducing Keypoint Action Tokens. 🤖 We translate visual observations and robot actions into a "language" that off-the-shelf LLMs can ingest and output. This transforms LLMs into *in-context, low-level imitation learning machines*. 🚀 Let me explain. 👇🧵

✨ Introducing Keypoint Action Tokens. 🤖 We translate visual observations and robot actions into a "language" that off-the-shelf LLMs can ingest and output. This transforms LLMs into in-context, low-level imitation learning machines. 🚀 Let me explain. 👇🧵

Norman Di Palo

23,115 Aufrufe • vor 2 Jahren

We can teach LLMs to write better robot code through natural language feedback. But can LLMs remember what they were taught and improve their teachability over time? Introducing our latest work, Learning to Learn Faster from Human Feedback with Language Model Predictive Control

We can teach LLMs to write better robot code through natural language feedback. But can LLMs remember what they were taught and improve their teachability over time? Introducing our latest work, Learning to Learn Faster from Human Feedback with Language Model Predictive Control

Jacky Liang

86,821 Aufrufe • vor 2 Jahren

NeurIPS 2025 Paper: LLMs are Reinforcement Learners 🤯! Surprisingly, we show that LLMs can solve RL tasks without any external component! We introduce Prompted Policy Search (ProPS), an RL method based only LLMs and in-context learning. [Paper]

NeurIPS 2025 Paper: LLMs are Reinforcement Learners 🤯! Surprisingly, we show that LLMs can solve RL tasks without any external component! We introduce Prompted Policy Search (ProPS), an RL method based only LLMs and in-context learning. [Paper]

Heni Ben Amor

51,248 Aufrufe • vor 11 Monaten

The robot is learning several novel tasks instantly, after just ONE demonstration each... Instant Policy makes it possible: no extra training, no weight updates, just pure in-context learning. It just got accepted at ICLR 2025, and it’s changing how robots learn. With just a single demo, a robot can pick up a new task and start performing it right away. Why this is a big deal: ✅ Learns tasks instantly with just one or a few demonstrations ✅ Improves over time as more demonstrations are given ✅ Uses simulation-based training with “pseudo-demonstrations” for scalability ✅ Can transfer skills across different robots and even follow language-defined tasks It brings in-context learning to robotics, opening up new possibilities for flexible, real-world automation. You can try it yourself: code and weights are available at • • • • Thank you to Edward Johns, Director of the Robot Learning Lab at Imperial College for sharing their work! 🙏

The robot is learning several novel tasks instantly, after just ONE demonstration each... Instant Policy makes it possible: no extra training, no weight updates, just pure in-context learning. It just got accepted at ICLR 2025, and it’s changing how robots learn. With just a single demo, a robot can pick up a new task and start performing it right away. Why this is a big deal: ✅ Learns tasks instantly with just one or a few demonstrations ✅ Improves over time as more demonstrations are given ✅ Uses simulation-based training with “pseudo-demonstrations” for scalability ✅ Can transfer skills across different robots and even follow language-defined tasks It brings in-context learning to robotics, opening up new possibilities for flexible, real-world automation. You can try it yourself: code and weights are available at • • • • Thank you to Edward Johns, Director of the Robot Learning Lab at Imperial College for sharing their work! 🙏

Ilir Aliu

46,285 Aufrufe • vor 1 Jahr

Robot Learning needs 4D world models! Robot Learning needs 4D world models! Robot Learning needs 4D world models! We introduce TesserAct, a 4D embodied world model that can simulate how agents interact with the 3D world over time! We achieve this by simply extending a pre-trained 2D video generation model to jointly predict RGB, depth, and surface normals. It enables: 1️⃣ Much better policy learning in the wild 2️⃣ Temporal + spatial coherence in 4D dynamic prediction 3️⃣ Novel view synthesis for embodied scenes Code: Paper Link: Project page:

Robot Learning needs 4D world models! Robot Learning needs 4D world models! Robot Learning needs 4D world models! We introduce TesserAct, a 4D embodied world model that can simulate how agents interact with the 3D world over time! We achieve this by simply extending a pre-trained 2D video generation model to jointly predict RGB, depth, and surface normals. It enables: 1️⃣ Much better policy learning in the wild 2️⃣ Temporal + spatial coherence in 4D dynamic prediction 3️⃣ Novel view synthesis for embodied scenes Code: Paper Link: Project page:

Chuang Gan

43,265 Aufrufe • vor 1 Jahr

This is a single uncut video, showing a robot learning several tasks instantly, after just one demonstration each ... This is possible because we've now been able to achieve in-context learning for everyday robotics tasks, and I'm very excited to announce our latest paper: 🎆 Instant Policy: In-Context Imitation Learning via Graph Diffusion 🎆 (1/6) 🧵👇

This is a single uncut video, showing a robot learning several tasks instantly, after just one demonstration each ... This is possible because we've now been able to achieve in-context learning for everyday robotics tasks, and I'm very excited to announce our latest paper: 🎆 Instant Policy: In-Context Imitation Learning via Graph Diffusion 🎆 (1/6) 🧵👇

Edward Johns

74,683 Aufrufe • vor 1 Jahr

Introducing Adapt-1 Machina: Learning Continuous Control Sequences With Adapt-1 Machina, we explore how continuous sequences of machine controls can be learned through reinforcement learning, starting with coarse outcome feedback and without demonstrations or a critic.

Introducing Adapt-1 Machina: Learning Continuous Control Sequences With Adapt-1 Machina, we explore how continuous sequences of machine controls can be learned through reinforcement learning, starting with coarse outcome feedback and without demonstrations or a critic.

Rei

42,587 Aufrufe • vor 1 Monat

What can we learn from demonstrations of long-horizon tasks? I am presenting our #CoRL2024 paper "Learning Compositional Behaviors from Demonstration and Language" today, showing we can learn a library of behaviors that can be composed to solve new tasks.

What can we learn from demonstrations of long-horizon tasks? I am presenting our #CoRL2024 paper "Learning Compositional Behaviors from Demonstration and Language" today, showing we can learn a library of behaviors that can be composed to solve new tasks.

Weiyu Liu

10,091 Aufrufe • vor 1 Jahr

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Sakana AI

179,465 Aufrufe • vor 1 Jahr

LLMs deployed in specialized domains, from enterprise codebases to law firms, dig through a mountain of unstructured context! Our new paper, Librarian Agents, asks: can an agent adapt by curating its own long-context workspace, anticipating queries before they arrive? 🧵⬇️

LLMs deployed in specialized domains, from enterprise codebases to law firms, dig through a mountain of unstructured context! Our new paper, Librarian Agents, asks: can an agent adapt by curating its own long-context workspace, anticipating queries before they arrive? 🧵⬇️

Anikait Singh

43,470 Aufrufe • vor 7 Tagen

We present ClutterGen that tackles two important research problems in robotics: 1) how to generate physically compliant cluttered scene for robot learning? 2) how to stably place objects instead of the current "pick and drop" paradigm? Our key idea is to formulate physically compliant cluttered scene generation as a sequential object placement problem, such that we can look at scene generation as a reinforcement learning problem. The learned environment generator can be easily applied to solve truly "pick and place" tasks. Our robot can figure out how to place an object in a stable manner into a cluttered environment. Please check out our project website with video, code, and paper to learn more! Great work done w/ Yinsen Jia at Duke General Robotics Lab (

We present ClutterGen that tackles two important research problems in robotics: 1) how to generate physically compliant cluttered scene for robot learning? 2) how to stably place objects instead of the current "pick and drop" paradigm? Our key idea is to formulate physically compliant cluttered scene generation as a sequential object placement problem, such that we can look at scene generation as a reinforcement learning problem. The learned environment generator can be easily applied to solve truly "pick and place" tasks. Our robot can figure out how to place an object in a stable manner into a cluttered environment. Please check out our project website with video, code, and paper to learn more! Great work done w/ Yinsen Jia at Duke General Robotics Lab (

Boyuan Chen

20,715 Aufrufe • vor 2 Jahren

A team tested Pi0, Pi0 Fast, Gr00t, and ACT on real robot arms in manufacturing tasks. (🔖 Bookmark this for later!) The task was precise: place thin rectangular frames from a messy stack into a holder. The team fine-tuned each model on 100 real trajectories and compared training time, inference speed, motion quality, and success rates. ⬇️ Here’s a breakdown of what they found Pi0 (Original) ✅ Strongest overall performance in precise pick-and-place ✅ High success rate even in edge cases ✅ Longest training time (~11 hours, ~$30 per run) ✅ Inference time of 80 ms causes short pauses between actions Despite delays, it handles complex scenarios well… solid for high-precision tasks, but slow to train. Gr00t ✅ Trains fast (~2 hours, ~$5 per run) ✅ Performs almost as well as Pi0 on large-object tasks ✅ Struggles with fine precision; random movement in some trials ✅ More training didn’t fix jitter or random offsets Best suited for tasks where exact precision isn’t critical. Not ready for manufacturing-grade accuracy without more tuning. Pi0 Fast ✅ Promised faster training, but results were underwhelming ✅ Training at 6 hours still showed low success rates ✅ Inference was slower than expected ✅ Not reliable for generalizing even slightly new tasks Currently too unstable for real-world deployment. Doesn’t live up to the “Fast” name yet. ACT (Baseline) ✅ 200MB model—lightweight, but limited ✅ Struggles with stacked objects or ambiguous scenes ✅ Success rates around 70% in best-case setups ✅ Can’t match newer models on precision or generalization Still a solid baseline, but clearly a generation behind in robustness. 🚨 Extra Notes All newer models share a common issue: •Inference takes longer than a frame (80 ms vs 33 ms), so robots “pause” between chunks. •This results in jittery movements, but not a dealbreaker unless tasks are time-sensitive. Language-conditioned tasks also fell short: after training on two labeled tasks, the model couldn’t generalize to a third unseen combination using only text prompts. ✅ The good news? These models adapt well to new robot arms with quick fine-tuning. ❌ The bad news? There’s still no plug-and-play solution for improving performance after deployment. Reinforcement learning or DAgger-style data collection during real-world operation may be the next big step, something many teams in robotics are actively working on.

A team tested Pi0, Pi0 Fast, Gr00t, and ACT on real robot arms in manufacturing tasks. (🔖 Bookmark this for later!) The task was precise: place thin rectangular frames from a messy stack into a holder. The team fine-tuned each model on 100 real trajectories and compared training time, inference speed, motion quality, and success rates. ⬇️ Here’s a breakdown of what they found Pi0 (Original) ✅ Strongest overall performance in precise pick-and-place ✅ High success rate even in edge cases ✅ Longest training time (~11 hours, ~$30 per run) ✅ Inference time of 80 ms causes short pauses between actions Despite delays, it handles complex scenarios well… solid for high-precision tasks, but slow to train. Gr00t ✅ Trains fast (~2 hours, ~$5 per run) ✅ Performs almost as well as Pi0 on large-object tasks ✅ Struggles with fine precision; random movement in some trials ✅ More training didn’t fix jitter or random offsets Best suited for tasks where exact precision isn’t critical. Not ready for manufacturing-grade accuracy without more tuning. Pi0 Fast ✅ Promised faster training, but results were underwhelming ✅ Training at 6 hours still showed low success rates ✅ Inference was slower than expected ✅ Not reliable for generalizing even slightly new tasks Currently too unstable for real-world deployment. Doesn’t live up to the “Fast” name yet. ACT (Baseline) ✅ 200MB model—lightweight, but limited ✅ Struggles with stacked objects or ambiguous scenes ✅ Success rates around 70% in best-case setups ✅ Can’t match newer models on precision or generalization Still a solid baseline, but clearly a generation behind in robustness. 🚨 Extra Notes All newer models share a common issue: •Inference takes longer than a frame (80 ms vs 33 ms), so robots “pause” between chunks. •This results in jittery movements, but not a dealbreaker unless tasks are time-sensitive. Language-conditioned tasks also fell short: after training on two labeled tasks, the model couldn’t generalize to a third unseen combination using only text prompts. ✅ The good news? These models adapt well to new robot arms with quick fine-tuning. ❌ The bad news? There’s still no plug-and-play solution for improving performance after deployment. Reinforcement learning or DAgger-style data collection during real-world operation may be the next big step, something many teams in robotics are actively working on.

Ilir Aliu

21,844 Aufrufe • vor 1 Jahr

🚀🚀🚀 Ever wondered what it takes for robots to handle real-world household tasks? long-horizon execution, deformable object dexterity, and unseen object generalization — meet GR-3, ByteDance Seed’s new Vision-Language-Action (VLA) model! GR-3 is a generalizable Vision-Language-Action (VLA) model with strong capabilities in complex long-horizon tasks. It understands unseen abstract concepts, manipulates deformable objects robustly, and adapts to novel settings with minimal human data. ✨ Generalization: Generalizes well to unseen objects, environments, and even instructions with abstract concepts. ✨ Long-Horizon Manipulation: Completes long-horizon tasks with strong instruction-following capabilities. ✨ Deformable Object Manipulation: Manipulate deformable objects robustly. Project Page: Arxiv: #ByteDance #ByteDanceSeed #GR3 #VLA #Robotics #FoundationModels

🚀🚀🚀 Ever wondered what it takes for robots to handle real-world household tasks? long-horizon execution, deformable object dexterity, and unseen object generalization — meet GR-3, ByteDance Seed’s new Vision-Language-Action (VLA) model! GR-3 is a generalizable Vision-Language-Action (VLA) model with strong capabilities in complex long-horizon tasks. It understands unseen abstract concepts, manipulates deformable objects robustly, and adapts to novel settings with minimal human data. ✨ Generalization: Generalizes well to unseen objects, environments, and even instructions with abstract concepts. ✨ Long-Horizon Manipulation: Completes long-horizon tasks with strong instruction-following capabilities. ✨ Deformable Object Manipulation: Manipulate deformable objects robustly. Project Page: Arxiv: #ByteDance #ByteDanceSeed #GR3 #VLA #Robotics #FoundationModels

Xiao Ma

46,464 Aufrufe • vor 1 Jahr

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.

AK

249,798 Aufrufe • vor 3 Jahren