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。

✨ 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

1,522 subscribers

23,097 次观看 • 2 年前 •via X (Twitter)

科学技术教育

Anya Rossi• Live Now

Private livecam show

11 条评论

Norman Di Palo 的头像

Norman Di Palo2 年前

We represent visual inputs as keypoint tokens, and robot poses as action tokens. Keypoints are extracted via DINO, projected in 3D, and represented as text. The gripper pose is represented as a triplet of 3D points in the same euclidean space as the keypoints. There's more👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

Recording a demonstration means storing keypoint tokens from the scene and the action tokens of the trajectory. The numbers are represented as text (tokens), and can therefore go into an LLM prompt. This is where things get interesting... 👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

With the Keypoint Action Tokens "language", we unlock the pattern extraction and completion abilities of LLMs, from language to robotics. Given keypoint tokens from a new scene, we demonstrate that LLMs can output an effective trajectory of action tokens! 👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

Our experiments demonstrate that: 1) KAT surpasses SOTA methods like Diffusion Policies (DP) 2) the KAT representation can boost DP performance as well, improving on the end-to-end version KAT is also robust to distractors, as we show in the third execution of the video. 👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

A fascinating result is that KAT's performance improves as the underlying LLMs improve. Using GPT-4 Turbo outperforms GPT-3.5 Turbo. This suggests that the improvement and scaling of LLMs can lead to free improvement to robotics, using LLMs as imitation learning machines. 👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

In-context learning also means *zero training time*, and the ability to deploy and test skills immediately, or store them as text and retrieve them later. All our pipeline relies on pre-trained Foundation Models and can be deployed immediately. 👇🧵

Norman Di Palo 的头像

Norman Di Palo2 年前

Therefore, we show one of the most interesting uses of LLMs in robotics is *without any natural language*. Surprising! 💡 Supervised by @Ed__Johns. Paper: Website with videos:

XXXin 的头像

XXXin1 年前

Does it mean the user gives no prompt but the debit will determine which task to perform based on the initial scene configuration? What comes next?

Norman Di Palo 的头像

Norman Di Palo1 年前

This method was trained on single tasks in isolation. Extending it to receive language commands has been done eg here

ऋषिक तिवारी (Rishik Tiwari) 的头像

ऋषिक तिवारी (Rishik Tiwari)2 年前

Really interesting, I lately have been trying to do something similar but with egocentric cam and affordance maps (inspired from VoxPoser). While KAT may work for topdown RGBD, ego views will be difficult as it gets obstructed after grasping the obj.

Norman Di Palo 的头像

Norman Di Palo2 年前

Interesting! I worked quite a bit with wrist-cams and they offer advantages and disadvantages. I am currently working on some wrist-cam based extensions.

相关视频

Very excited to announce: Keypoint Action Tokens! We found that LLMs can be repurposed as "imitation learning engines" for robots, by representing both observations & actions as 3D keypoints, and feeding into an LLM for in-context learning. See: More 👇

Very excited to announce: Keypoint Action Tokens! We found that LLMs can be repurposed as "imitation learning engines" for robots, by representing both observations & actions as 3D keypoints, and feeding into an LLM for in-context learning. See: More 👇

Edward Johns

32,582 次观看 • 2 年前

We open-source Action Images — a new representation that translates 7-DoF robot actions into interpretable images. Video models are emerging as powerful robotic foundation models, but a key challenge remains: how can we seamlessly integrate robot policies into video models? Instead of representing actions as low-dimensional control tokens, Action Images provide a pixel-grounded action representation, reframing policy learning as a visual tracking problem! By unifying observations and actions in the same video space, Action Images enable a unified robotics world model that supports video-action joint generation, action-conditioned video generation, and action labeling! Code: Paper:

We open-source Action Images — a new representation that translates 7-DoF robot actions into interpretable images. Video models are emerging as powerful robotic foundation models, but a key challenge remains: how can we seamlessly integrate robot policies into video models? Instead of representing actions as low-dimensional control tokens, Action Images provide a pixel-grounded action representation, reframing policy learning as a visual tracking problem! By unifying observations and actions in the same video space, Action Images enable a unified robotics world model that supports video-action joint generation, action-conditioned video generation, and action labeling! Code: Paper:

Chuang Gan

20,281 次观看 • 2 个月前

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), our Conference on Robot Learning (CoRL) 2025 paper. Our RICL-pi0 model can adapt to unseen objects, novel motions, and new scenes with just ICL and RAG (retrieval-augmented generation). RICL-pi0 also boosts performance on the long-tail of tasks. A quick 1 minute video summary:

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), our Conference on Robot Learning (CoRL) 2025 paper. Our RICL-pi0 model can adapt to unseen objects, novel motions, and new scenes with just ICL and RAG (retrieval-augmented generation). RICL-pi0 also boosts performance on the long-tail of tasks. A quick 1 minute video summary:

Kaustubh Sridhar

52,158 次观看 • 1 年前

With OpenAI, Figure 01 can now have full conversations with people -OpenAI models provide high-level visual and language intelligence -Figure neural networks deliver fast, low-level, dexterous robot actions Everything in this video is a neural network:

With OpenAI, Figure 01 can now have full conversations with people -OpenAI models provide high-level visual and language intelligence -Figure neural networks deliver fast, low-level, dexterous robot actions Everything in this video is a neural network:

Figure

5,104,353 次观看 • 2 年前

What if robots, humans, and video models could all describe action in the same language? Introducing Hydra-0🐍: a generalist world model that represents robot actions as action flow. One interface. Many embodiments, tasks, and environments. 🧵1/7

What if robots, humans, and video models could all describe action in the same language? Introducing Hydra-0🐍: a generalist world model that represents robot actions as action flow. One interface. Many embodiments, tasks, and environments. 🧵1/7

Hongyu Li

53,116 次观看 • 21 天前

Yann LeCun says language isn’t intelligence. Predicting text doesn’t mean understanding reality. The real world is messy, physical, and causal and today’s LLMs barely touch that. The next leap is Physical AI: world models, cause and effect, real planning. Do you think LLMs can evolve into this, or do we need a completely new architecture?

Yann LeCun says language isn’t intelligence. Predicting text doesn’t mean understanding reality. The real world is messy, physical, and causal and today’s LLMs barely touch that. The next leap is Physical AI: world models, cause and effect, real planning. Do you think LLMs can evolve into this, or do we need a completely new architecture?

VraserX e/acc

76,220 次观看 • 7 个月前

Your bimanual manipulators might need a Robot Neck 🤖🦒 Introducing Vision in Action: Learning Active Perception from Human Demonstrations ViA learns task-specific, active perceptual strategies—such as searching, tracking, and focusing—directly from human demos, enabling robust visuomotor policies under visual occlusions. 🧵👇

Your bimanual manipulators might need a Robot Neck 🤖🦒 Introducing Vision in Action: Learning Active Perception from Human Demonstrations ViA learns task-specific, active perceptual strategies—such as searching, tracking, and focusing—directly from human demos, enabling robust visuomotor policies under visual occlusions. 🧵👇

Haoyu Xiong

123,959 次观看 • 1 年前

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 次观看 • 3 年前

Achieve SOTA on multimodal, complex tasks with minimal supervision! Introducing In-Context Abstraction Learning (ICAL): a method that extracts insights from visual demos and human feedback to continually teach multimodal LLM agents 🚀🤖🧠 ICAL learns from sub-optimal demonstrations and human feedback by transforming noisy trajectories into optimized, annotated examples that improve multimodal decision-making in new domains. ICAL is the first to show state-of-the-art performance across three challenging domains: we obtain a 34% relative improvement on TEACh household robot instruction following from dialogue, a 59% relative improvement on Visual Web Arena for autonomous visual web agents, and beat few-shot GPT-4V and supervised baselines on Ego4D action forecasting.

Achieve SOTA on multimodal, complex tasks with minimal supervision! Introducing In-Context Abstraction Learning (ICAL): a method that extracts insights from visual demos and human feedback to continually teach multimodal LLM agents 🚀🤖🧠 ICAL learns from sub-optimal demonstrations and human feedback by transforming noisy trajectories into optimized, annotated examples that improve multimodal decision-making in new domains. ICAL is the first to show state-of-the-art performance across three challenging domains: we obtain a 34% relative improvement on TEACh household robot instruction following from dialogue, a 59% relative improvement on Visual Web Arena for autonomous visual web agents, and beat few-shot GPT-4V and supervised baselines on Ego4D action forecasting.

Gabriel Sarch

19,340 次观看 • 2 年前

Andrej Karpathy beautifully explains the fundamental difference of learning between a human and an LLM. > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” many of the best minds in this field think LLMs are not really learning anything, and therefore are incapable of surpassing human-level intelligence.

Andrej Karpathy beautifully explains the fundamental difference of learning between a human and an LLM. > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” many of the best minds in this field think LLMs are not really learning anything, and therefore are incapable of surpassing human-level intelligence.

ℏεsam

895,349 次观看 • 10 个月前

Muse Spark 1.2 supports a broad range of multimodal tasks, from turning visuals into working code to translating perception into physical action. It also brings robust audio-visual understanding to enable video-heavy workflows common in real-world enterprise use. Today, we’re sharing new evals and demos that illustrate the breadth of the model’s visual understanding and reasoning capabilities. Let’s start with a demo that shows how Muse Spark parses multimodal observations and calls tools to guide a robot to navigate in an unstructured environment to find a rubber duck. 🧵👇

Muse Spark 1.2 supports a broad range of multimodal tasks, from turning visuals into working code to translating perception into physical action. It also brings robust audio-visual understanding to enable video-heavy workflows common in real-world enterprise use. Today, we’re sharing new evals and demos that illustrate the breadth of the model’s visual understanding and reasoning capabilities. Let’s start with a demo that shows how Muse Spark parses multimodal observations and calls tools to guide a robot to navigate in an unstructured environment to find a rubber duck. 🧵👇

AI at Meta

54,184 次观看 • 22 天前

Turn any GitHub repository into LLM-ready text! Simply replace "hub" with "ingest" in a GitHub URL and receive a prompt-friendly text ingest for LLMs. Gitingest is 100% open-source and provides: - Directory structure - A brief summary of the project - The entire content as LLM-ready text Plus, it comes with a nice python package and you can run the UI locally! Stay tuned, I'm working on something really cool with this!✨ Link to the GitHub repo in the next tweet! _____ Find me → Akshay 🚀 ✔️ For more insights and tutorials on ML and AI Engineering!

Turn any GitHub repository into LLM-ready text! Simply replace "hub" with "ingest" in a GitHub URL and receive a prompt-friendly text ingest for LLMs. Gitingest is 100% open-source and provides: - Directory structure - A brief summary of the project - The entire content as LLM-ready text Plus, it comes with a nice python package and you can run the UI locally! Stay tuned, I'm working on something really cool with this!✨ Link to the GitHub repo in the next tweet! _____ Find me → Akshay 🚀 ✔️ For more insights and tutorials on ML and AI Engineering!

Akshay 🚀

191,449 次观看 • 1 年前

In the era of long-context LLMs, I needed a way to quickly convert a GitHub repo into a text file to feed as a prompt. Introducing RepoToTextForLLMs 📖 Not only does it return the full repo in one file, but it also appends a super prompt for analysis and understanding 👇

In the era of long-context LLMs, I needed a way to quickly convert a GitHub repo into a text file to feed as a prompt. Introducing RepoToTextForLLMs 📖 Not only does it return the full repo in one file, but it also appends a super prompt for analysis and understanding 👇

Pietro Schirano

104,922 次观看 • 2 年前

Andrej Karpathy (Anthropic Team) dropped a full 7-hour course on "Intro and Deep Dive to LLMs and Tokenizer: • 00:00 - Intro to Large Language Models • 00:59:28 - Deep dive into LLMs (Karpathy method) • 03:31:43 - Let s build the GPT Tokenizer This Karpathy guide can replace a $100K Yale LLM senior degree. Start watching right now, then read article below

Andrej Karpathy (Anthropic Team) dropped a full 7-hour course on "Intro and Deep Dive to LLMs and Tokenizer: • 00:00 - Intro to Large Language Models • 00:59:28 - Deep dive into LLMs (Karpathy method) • 03:31:43 - Let s build the GPT Tokenizer This Karpathy guide can replace a $100K Yale LLM senior degree. Start watching right now, then read article below

Morty

44,481 次观看 • 28 天前

.Richard Sutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI

.Richard Sutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training phase - the agent just learns on-the-fly - like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete. I did my best to represent the view that LLMs will function as the foundation on which this experiential learning can happen. Some sparks flew. 0:00:00 – Are LLMs a dead-end? 0:13:51 – Do humans do imitation learning? 0:23:57 – The Era of Experience 0:34:25 – Current architectures generalize poorly out of distribution 0:42:17 – Surprises in the AI field 0:47:28 – Will The Bitter Lesson still apply after AGI? 0:54:35 – Succession to AI

Dwarkesh Patel

3,087,972 次观看 • 11 个月前

AI web agents like Operator and Anthropic’s Computer Use can operate a browser, but the LLMs inside are brittle, and you can’t trust what’s on the web. In this 🧵, I’ll show how adversaries can fool Anthropic’s web agent into sending phishing emails or revealing credit card info.

AI web agents like Operator and Anthropic’s Computer Use can operate a browser, but the LLMs inside are brittle, and you can’t trust what’s on the web. In this 🧵, I’ll show how adversaries can fool Anthropic’s web agent into sending phishing emails or revealing credit card info.

Micah Goldblum

42,969 次观看 • 1 年前

Announcing How Transformer LLMs Work, created with Jay Alammar and Maarten Grootendorst, co-authors of the beautifully illustrated book, “Hands-On Large Language Models.” This course offers a deep dive into the inner workings of the transformer architecture that powers large language models (LLMs). The transformer architecture revolutionized generative AI; in fact, the "GPT" in ChatGPT stands for "Generative Pre-Trained Transformer." Originally introduced in the Google Brain team's groundbreaking 2017 paper "Attention Is All You Need," by Vaswani and others, transformers were a highly scalable model for machine translation tasks. Variants of this architecture now power today’s LLMs such as those from OpenAI, Google, Meta, Cohere, Anthropic and DeepSeek. In this course, you’ll learn in detail how LLMs process text. You'll also work through code examples that illustrate that transformer's individual components. In details, you’ll learn: - How the representation of language has evolved, from Bag-of-Words to Word2Vec embeddings to the transformer architecture that captures a word's meanings taking into account the context of other words in the input. - How inputs are broken down into tokens before they are sent to the language model. - The details of a transformer's main stages: Tokenization and embedding, the stack of transformer blocks, and the language model head. - The inner workings of the transformer block, including attention, which calculates relevance scores, and the feedforward layer, which incorporates stored information learned in training. - How cached calculations make transformers faster. - Some of the most recent ideas in the latest models such as Mixture-of-Experts (MoE) which uses multiple sub-models and a router on each layer to improve the quality of LLMs. By the end of this course, you’ll have a deep understanding of how LLMs actually process text and be able to read through papers describing the latest models and understand the details. Gaining this intuition will improve your approach to building LLM applications. Please sign up here:

Announcing How Transformer LLMs Work, created with Jay Alammar and Maarten Grootendorst, co-authors of the beautifully illustrated book, “Hands-On Large Language Models.” This course offers a deep dive into the inner workings of the transformer architecture that powers large language models (LLMs). The transformer architecture revolutionized generative AI; in fact, the "GPT" in ChatGPT stands for "Generative Pre-Trained Transformer." Originally introduced in the Google Brain team's groundbreaking 2017 paper "Attention Is All You Need," by Vaswani and others, transformers were a highly scalable model for machine translation tasks. Variants of this architecture now power today’s LLMs such as those from OpenAI, Google, Meta, Cohere, Anthropic and DeepSeek. In this course, you’ll learn in detail how LLMs process text. You'll also work through code examples that illustrate that transformer's individual components. In details, you’ll learn: - How the representation of language has evolved, from Bag-of-Words to Word2Vec embeddings to the transformer architecture that captures a word's meanings taking into account the context of other words in the input. - How inputs are broken down into tokens before they are sent to the language model. - The details of a transformer's main stages: Tokenization and embedding, the stack of transformer blocks, and the language model head. - The inner workings of the transformer block, including attention, which calculates relevance scores, and the feedforward layer, which incorporates stored information learned in training. - How cached calculations make transformers faster. - Some of the most recent ideas in the latest models such as Mixture-of-Experts (MoE) which uses multiple sub-models and a router on each layer to improve the quality of LLMs. By the end of this course, you’ll have a deep understanding of how LLMs actually process text and be able to read through papers describing the latest models and understand the details. Gaining this intuition will improve your approach to building LLM applications. Please sign up here:

Andrew Ng

259,920 次观看 • 1 年前