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.

Introducing Scalable Option Learning (SOL☀️), a blazingly fast hierarchical RL algorithm that makes progress on long-horizon tasks and demonstrates positive scaling trends on the largely unsolved NetHack benchmark, when trained for 30 billion samples. Details, paper and code in >

Mikael Henaff

1,907 subscribers

21,956 Aufrufe • vor 1 Jahr •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

5 Kommentare

Profilbild von Mikael Henaff

Mikael Henaffvor 1 Jahr

Hierarchy is a natural way to tackle long horizons, but until now has remained at relatively small scale. With SOL, we identify and solve bottlenecks in scaling hierarchical RL, resulting in a ~35-580x speed increase over prior hierarchical methods.

Profilbild von Mikael Henaff

Mikael Henaffvor 1 Jahr

We demonstrate SOL's performance and scalability by training hierarchical agents for 30B steps on the complex game of NetHack, significantly outperforming flat agents and demonstrating promising scaling trends. Our agents still seem to be improving, even after 30B steps.

Profilbild von Mikael Henaff

Mikael Henaffvor 1 Jahr

SOL can be run on any RL problem for which we can define a few reasonable intrinsic rewards. We include some simple PointMaze and MiniHack environments to show this generality - this may also be useful for others working in HRL since they are faster to iterate on than NetHack.

Profilbild von Mikael Henaff

Mikael Henaffvor 1 Jahr

All details are in our paper and code release: It was lots of fun working with Scott Fujimoto, @mitrma and Mike Rabbat on this!

Profilbild von Robert Joseph

Robert Josephvor 5 Monaten

This is so cool! Thanks: ) been looking for good HRL repos!

Ähnliche Videos

Introducing Adjoint Sampling, a new learning algorithm that trains generative models based on scalar rewards. Based on theoretical foundations developed by FAIR, Adjoint Sampling leads to a highly scalable practical algorithm, and can become the foundation for further research into highly scalable sampling methods. Read our research paper on Adjoint Sampling and download the model, code, and benchmark ➡️

Introducing Adjoint Sampling, a new learning algorithm that trains generative models based on scalar rewards. Based on theoretical foundations developed by FAIR, Adjoint Sampling leads to a highly scalable practical algorithm, and can become the foundation for further research into highly scalable sampling methods. Read our research paper on Adjoint Sampling and download the model, code, and benchmark ➡️

AI at Meta

36,987 Aufrufe • vor 1 Jahr

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: 💻 Code: 🗣️ Discord: 🤗 Data & Models: More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: 💻 Code: 🗣️ Discord: 🤗 Data & Models: More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵

Jacky Kwok

1,054,183 Aufrufe • vor 10 Tagen

Additionally, looking towards the future, we’re releasing PARTNR: a benchmark for Planning And Reasoning Tasks in humaN-Robot collaboration. Built on Habitat 3.0, it’s the largest benchmark of its kind to study and evaluate human-robot collaboration in household activities By providing a standardized benchmark and dataset we hope to enable new research on robots that can not only operate in isolation, but in collaboration with people. Details and code ➡️

Additionally, looking towards the future, we’re releasing PARTNR: a benchmark for Planning And Reasoning Tasks in humaN-Robot collaboration. Built on Habitat 3.0, it’s the largest benchmark of its kind to study and evaluate human-robot collaboration in household activities By providing a standardized benchmark and dataset we hope to enable new research on robots that can not only operate in isolation, but in collaboration with people. Details and code ➡️

AI at Meta

20,244 Aufrufe • vor 1 Jahr

How well do today’s frontier models handle long-horizon, multi-step web agent tasks, such as identifying the top 25 U.S. CS PhD programs with ML/AI faculty likely accepting students and compiling the results into a structured sheet? Check out our new work on Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks Paper: Leaderboard: We introduce Odysseys, a benchmark of 200 long-horizon tasks derived from real browsing sessions and evaluated on the live Internet. We show that binary pass/fail is inadequate in this setting and propose rubric-based evaluation, which better aligns with human judgment and provides more informative signals. Across leading models, the best achieves only 44.5% success, leaving substantial headroom. We further introduce a Trajectory Efficiency metric (rubric score per step) and find efficiency remains extremely low (1.15%), highlighting a key bottleneck. Odysseys provides a realistic benchmark for measuring progress toward web agents capable of sustained, efficient, real-world operation. See a more detailed thread by Jing Yu Koh.

How well do today’s frontier models handle long-horizon, multi-step web agent tasks, such as identifying the top 25 U.S. CS PhD programs with ML/AI faculty likely accepting students and compiling the results into a structured sheet? Check out our new work on Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks Paper: Leaderboard: We introduce Odysseys, a benchmark of 200 long-horizon tasks derived from real browsing sessions and evaluated on the live Internet. We show that binary pass/fail is inadequate in this setting and propose rubric-based evaluation, which better aligns with human judgment and provides more informative signals. Across leading models, the best achieves only 44.5% success, leaving substantial headroom. We further introduce a Trajectory Efficiency metric (rubric score per step) and find efficiency remains extremely low (1.15%), highlighting a key bottleneck. Odysseys provides a realistic benchmark for measuring progress toward web agents capable of sustained, efficient, real-world operation. See a more detailed thread by Jing Yu Koh.

Russ Salakhutdinov

22,518 Aufrufe • vor 5 Monaten

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

179,423 Aufrufe • vor 1 Jahr

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more

Prime Intellect

1,150,419 Aufrufe • vor 10 Monaten

Attune (Attune) makes blazingly fast build tools for developers that require zero migration effort. Built by developers with a decade of experience in build systems and developer tools. Congrats on the launch, xin and Eliza Zhang!

Attune (Attune) makes blazingly fast build tools for developers that require zero migration effort. Built by developers with a decade of experience in build systems and developer tools. Congrats on the launch, xin and Eliza Zhang!

Y Combinator

16,787 Aufrufe • vor 1 Jahr

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

52,897 Aufrufe • vor 1 Jahr

A few weeks ago, we shared our progress on articulated objects and long-horizon tasks. Here are two representative examples: - We've been steadily expanding our asset library to cover more articulated objects. Articulated objects have always been a challenging asset class to handle in simulation. Interacting with them requires robots to master atomic skills such as pushing, pulling, opening, and closing, and to understand part structure, interaction constraints, and how the object moves. - Long-horizon tasks can now be generated at scale. Long-horizon tasks are the other hard category: they require chaining multiple sub-goals in sequence. A failure early in the task can cascade and make the rest unrecoverable. Axis is scaling along three dimensions at once: data volume, data quality, and task difficulty.

A few weeks ago, we shared our progress on articulated objects and long-horizon tasks. Here are two representative examples: - We've been steadily expanding our asset library to cover more articulated objects. Articulated objects have always been a challenging asset class to handle in simulation. Interacting with them requires robots to master atomic skills such as pushing, pulling, opening, and closing, and to understand part structure, interaction constraints, and how the object moves. - Long-horizon tasks can now be generated at scale. Long-horizon tasks are the other hard category: they require chaining multiple sub-goals in sequence. A failure early in the task can cascade and make the rest unrecoverable. Axis is scaling along three dimensions at once: data volume, data quality, and task difficulty.

Axis Robotics

14,217 Aufrufe • vor 3 Monaten

Building upon SimpleVLA-RL, we have implemented real-world RL on long-horizon dexterous tasks and witnessed a non-trivial (~relatively 300%) performance improvement over the SFT model, along with surprising capabilities on auto-recovery. Blog coming soon. The entire process uses very little data and training compute—basically costing no more than a single robotic arm—hinting that real-world generality for machines is actually within sight.

Building upon SimpleVLA-RL, we have implemented real-world RL on long-horizon dexterous tasks and witnessed a non-trivial (~relatively 300%) performance improvement over the SFT model, along with surprising capabilities on auto-recovery. Blog coming soon. The entire process uses very little data and training compute—basically costing no more than a single robotic arm—hinting that real-world generality for machines is actually within sight.

Ning Ding

93,483 Aufrufe • vor 9 Monaten

The power of generative models — now embodied in humanoids. Announcing DreamControl –– After a year-long research effort at General Robotics — we present a scalable framework for whole-body humanoid control that fuses diffusion priors with reinforcement learning to unlock real-world scene interaction. Diffusion + RL → natural whole-body skills on real robots. DreamControl enables humanoids to move beyond locomotion demos → performing natural, human-like skills such as –– Picking & lifting objects, Opening drawers & doors, Precise punching, kicking, and jumping, Bimanual manipulation tasks Our key innovation: a diffusion prior over human motion that guides RL, eliminating the need for massive teleoperation datasets, and producing motions that look human while transferring to real hardware. Trained purely in simulation, deployed on the Unitree G1 humanoid, DreamControl policies run in real time, bridging sim-to-real with unprecedented naturalness. We leverage a novel hybrid edge + cloud infrastructure that runs RL-trained policies on the edge backed by powerful AI models running in the cloud This is the next step in General Robotics’ journey toward general-purpose humanoid assistants that interact, adapt, and assist autonomously. Paper: Blog: 1/n

The power of generative models — now embodied in humanoids. Announcing DreamControl –– After a year-long research effort at General Robotics — we present a scalable framework for whole-body humanoid control that fuses diffusion priors with reinforcement learning to unlock real-world scene interaction. Diffusion + RL → natural whole-body skills on real robots. DreamControl enables humanoids to move beyond locomotion demos → performing natural, human-like skills such as –– Picking & lifting objects, Opening drawers & doors, Precise punching, kicking, and jumping, Bimanual manipulation tasks Our key innovation: a diffusion prior over human motion that guides RL, eliminating the need for massive teleoperation datasets, and producing motions that look human while transferring to real hardware. Trained purely in simulation, deployed on the Unitree G1 humanoid, DreamControl policies run in real time, bridging sim-to-real with unprecedented naturalness. We leverage a novel hybrid edge + cloud infrastructure that runs RL-trained policies on the edge backed by powerful AI models running in the cloud This is the next step in General Robotics’ journey toward general-purpose humanoid assistants that interact, adapt, and assist autonomously. Paper: Blog: 1/n

Ashish Kapoor

118,521 Aufrufe • vor 1 Jahr

We are releasing 4M-21 with a permissive license, including its source code and trained models. It's a pretty effective multimodal model that solves 10s of tasks & modalities. See the demo code, sample results, and the tokenizers of diverse modalities on the website. IMO, the multitask learning aspect of multimodal models has really taken a step forward. We can train a single model on many diverse tasks with ~SOTA accuracy. But a long way to go in terms of transfer/emergence. 🌐 ⌨️ Joint work w/ EPFL Apple.

We are releasing 4M-21 with a permissive license, including its source code and trained models. It's a pretty effective multimodal model that solves 10s of tasks & modalities. See the demo code, sample results, and the tokenizers of diverse modalities on the website. IMO, the multitask learning aspect of multimodal models has really taken a step forward. We can train a single model on many diverse tasks with ~SOTA accuracy. But a long way to go in terms of transfer/emergence. 🌐 ⌨️ Joint work w/ EPFL Apple.

Amir Zamir

69,604 Aufrufe • vor 2 Jahren

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵

Deyao Zhu

363,667 Aufrufe • vor 3 Monaten

Can reinforcement learning from AI feedback unlock new capabilities in AI agents? Introducing Motif, an LLM-powered method for intrinsic motivation from AI feedback. Motif extracts reward functions from Llama 2's preferences and uses them to train agents with reinforcement learning. On the complex NetHack game, Motif solves previously unsolved tasks without needing any expert demonstrations. Surprisingly, Motif's reward leads to better game score than the one obtained by using the score itself as a reward. Given access to an event captioning mechanism, a few properties make Motif a general method: • it is entirely based on open models • the LLM doesn't need direct access to the environment dynamics (e.g., its source code) • the LLM doesn't need to understand observation and action spaces The best part? You can start using Motif right now, even on a small compute budget: the whole pipeline can take less than two GPU-days. Feel free to read our paper and try our code out. Paper: Code: Blog post: Work co-lead by Martin Klissarov and myself, with Shagun Sodhani Roberta Raileanu Pierre-Luc Bacon Pascal Vincent Amy Zhang Mikael Henaff Learn more in the thread 🧵

Can reinforcement learning from AI feedback unlock new capabilities in AI agents? Introducing Motif, an LLM-powered method for intrinsic motivation from AI feedback. Motif extracts reward functions from Llama 2's preferences and uses them to train agents with reinforcement learning. On the complex NetHack game, Motif solves previously unsolved tasks without needing any expert demonstrations. Surprisingly, Motif's reward leads to better game score than the one obtained by using the score itself as a reward. Given access to an event captioning mechanism, a few properties make Motif a general method: • it is entirely based on open models • the LLM doesn't need direct access to the environment dynamics (e.g., its source code) • the LLM doesn't need to understand observation and action spaces The best part? You can start using Motif right now, even on a small compute budget: the whole pipeline can take less than two GPU-days. Feel free to read our paper and try our code out. Paper: Code: Blog post: Work co-lead by Martin Klissarov and myself, with Shagun Sodhani Roberta Raileanu Pierre-Luc Bacon Pascal Vincent Amy Zhang Mikael Henaff Learn more in the thread 🧵

Pierluca D'Oro

311,942 Aufrufe • vor 2 Jahren

Progress in open models is keeping Big AI labs up at night, and I'm here for it! We have a brand new open-weight multimodal model optimized for long-horizon tasks. This model is really good at something: it can work on tasks that keep evolving over time. • 280B total parameters, but only 16B active • 512K context window • Understands text, images, and audio • Strong reasoning, coding, and tool use But the best of all: the model learns and adapts to new information! Imagine you start running an agent today to solve a problem, and while it's working, you get new information that changes the initial conditions, or you change your mind. The agents you run today don't have issues with short tasks and goals that don't change, but reality is messy, and that makes it hard for long-horizon agents to succeed. The new dots3-note Preview model introduces TEMPO. TEMPO is a new reinforcement learning technique that lets the model periodically pause and critique its own progress. Basically, from time to time, the agent asks itself: "Am I getting closer to the goal, or am I wasting my time?" The same model switches between actor and critic. The actor works on the problem. The critic looks at the current state, reasons about how much progress it has made, and determines what should happen next. TEMPO gives the model feedback along the way. This is huge for any agent that can work on long-horizon tasks without wasting its time.

Santiago

80,792 Aufrufe • vor 1 Monat

Introducing LDA, a latent world action foundation model that, for the first time, unifies the utilization of heterogeneous embodied data across simulation and reality, humans and robots, and varying levels of action quality and annotation. By breaking long-standing data silos in embodied intelligence, LDA enables the field, much like GPT did for language, to benefit continuously from scaling data, marking the transition into a new era of scalable learning. #Galbot #Robotics #Innovation #AI #Technology #Humanoid #WorldModel

Introducing LDA, a latent world action foundation model that, for the first time, unifies the utilization of heterogeneous embodied data across simulation and reality, humans and robots, and varying levels of action quality and annotation. By breaking long-standing data silos in embodied intelligence, LDA enables the field, much like GPT did for language, to benefit continuously from scaling data, marking the transition into a new era of scalable learning. #Galbot #Robotics #Innovation #AI #Technology #Humanoid #WorldModel

Galbot

40,883 Aufrufe • vor 5 Monaten

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MTS

100,519 Aufrufe • vor 2 Monaten

Can vision-language-action (VLA) models generalize to diverse OOD tasks and align with customized objectives? 🤔 🚀 We introduce GRAPE, a plug-and-play algorithm to generalize robot policies via preference alignment. GRAPE unfolds three benefits to boost the generalizability of VLAs: 👉1. GRAPE aligns VLAs on a trajectory level and endows the model with the ability for global decision-making, instead of merely cloning behavior; 👉2. GRAPE implicitly models reward from both successful and failed trials to boost generalizability to diverse tasks; 👉3. GRAPE adopts a scalable preference synthesis algorithm to rank trajectories with preferences that align with arbitrary objectives. Our experiments on a diverse array of real-world and simulated robotic tasks reveal: 1⃣GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 60.36%; 2⃣GRAPE is versatile to be aligned with diverse objectives and reduce collision rates by 44.31% or rollout length by 11.15% when aligning towards safer or more efficient manipulation policy, respectively. Check out our full project for more details: 🔥 Paper: 🔥 Project: 🔥 Code:

Can vision-language-action (VLA) models generalize to diverse OOD tasks and align with customized objectives? 🤔 🚀 We introduce GRAPE, a plug-and-play algorithm to generalize robot policies via preference alignment. GRAPE unfolds three benefits to boost the generalizability of VLAs: 👉1. GRAPE aligns VLAs on a trajectory level and endows the model with the ability for global decision-making, instead of merely cloning behavior; 👉2. GRAPE implicitly models reward from both successful and failed trials to boost generalizability to diverse tasks; 👉3. GRAPE adopts a scalable preference synthesis algorithm to rank trajectories with preferences that align with arbitrary objectives. Our experiments on a diverse array of real-world and simulated robotic tasks reveal: 1⃣GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 60.36%; 2⃣GRAPE is versatile to be aligned with diverse objectives and reduce collision rates by 44.31% or rollout length by 11.15% when aligning towards safer or more efficient manipulation policy, respectively. Check out our full project for more details: 🔥 Paper: 🔥 Project: 🔥 Code:

Huaxiu Yao

20,071 Aufrufe • vor 1 Jahr

Can AI agents adapt zero-shot, to complex multi-step language instructions in open-ended environments? We present MaestroMotif, a method for AI-assisted skill design that produces highly capable and steerable hierarchical agents. To the best of our knowledge, it is the first method that, without expert labeled datasets, solves compositional tasks requiring hundreds of steps for completion. All the modules within MaestroMotif are learned from interaction: from the highest level of planning to the lowest-level of sensorimotor control. On the open-ended domain of NetHack, it surpasses existing approaches, including those that are fine-tuned specifically for each task. At the heart of MaestroMotif is the idea that decomposing a task into subtasks significantly helps decision making. MaestroMotif leverages an agent designer's intuition about a domain to identify important skills and describe them in natural language. These short descriptions then get converted into adaptable hierarchical agents through AI feedback and in-context learning. Our paper was recently published at ICLR 2025 and we open-source the whole project including the code, prompts and pre-trained models. Paper: Code: NotebookLM Podcast: This work was done with the amazing Mikael Henaff, Roberta Raileanu, Shagun Sodhani, Pascal Vincent, Amy Zhang, Pierre-Luc Bacon, Doina Precup, with equal supervision by Marlos C. Machado and Pierluca D'Oro. Take a look at the following thread:

Can AI agents adapt zero-shot, to complex multi-step language instructions in open-ended environments? We present MaestroMotif, a method for AI-assisted skill design that produces highly capable and steerable hierarchical agents. To the best of our knowledge, it is the first method that, without expert labeled datasets, solves compositional tasks requiring hundreds of steps for completion. All the modules within MaestroMotif are learned from interaction: from the highest level of planning to the lowest-level of sensorimotor control. On the open-ended domain of NetHack, it surpasses existing approaches, including those that are fine-tuned specifically for each task. At the heart of MaestroMotif is the idea that decomposing a task into subtasks significantly helps decision making. MaestroMotif leverages an agent designer's intuition about a domain to identify important skills and describe them in natural language. These short descriptions then get converted into adaptable hierarchical agents through AI feedback and in-context learning. Our paper was recently published at ICLR 2025 and we open-source the whole project including the code, prompts and pre-trained models. Paper: Code: NotebookLM Podcast: This work was done with the amazing Mikael Henaff, Roberta Raileanu, Shagun Sodhani, Pascal Vincent, Amy Zhang, Pierre-Luc Bacon, Doina Precup, with equal supervision by Marlos C. Machado and Pierluca D'Oro. Take a look at the following thread:

Martin Klissarov

80,307 Aufrufe • vor 1 Jahr

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has all the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Sonya Huang 🐥

81,332 Aufrufe • vor 4 Monaten