Loading video...

Video Failed to Load

Go Home

Robots might learn better from video than from language! 📼 Most Vision-Language-Action (VLA) models learn what to do from text, but still struggle with how things move in the real world. That makes them data-hungry and slow to train. mimic video takes a different route. Instead of grounding robot...

lukas_m_ziegler's profile picture

Lukas Ziegler

60,056 subscribers

50,090 views • 8 months ago •via X (Twitter)

Science & Technology
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Turning video into humanoid robot motion! 🤳🏼 Training humanoid robots needs huge amounts of motion data, but real-world capture doesn’t scale. Mocap is expensive, dangerous edge cases are rare, and you can’t ask humans to repeatedly fall or crash. Video2Robot tackles this by converting videos into physics-grounded humanoid simulations. Motion is generated to respect balance, inertia, ground contact, and joint limits, then directly retargeted to robot simulators. One prompt can generate a full humanoid motion sequence, including multi-agent interactions and failure cases like falls or collisions, scenarios that are hard or impossible to capture safely in the real world. The pipeline is model-agnostic and works with existing video generators, making it a practical way to scale data for robots. If robots are going to operate in the real world, they need to be trained on the failures too, not just the perfect demos. Here's the GitHub: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →
0:26

Turning video into humanoid robot motion! 🤳🏼 Training humanoid robots needs huge amounts of motion data, but real-world capture doesn’t scale. Mocap is expensive, dangerous edge cases are rare, and you can’t ask humans to repeatedly fall or crash. Video2Robot tackles this by converting videos into physics-grounded humanoid simulations. Motion is generated to respect balance, inertia, ground contact, and joint limits, then directly retargeted to robot simulators. One prompt can generate a full humanoid motion sequence, including multi-agent interactions and failure cases like falls or collisions, scenarios that are hard or impossible to capture safely in the real world. The pipeline is model-agnostic and works with existing video generators, making it a practical way to scale data for robots. If robots are going to operate in the real world, they need to be trained on the failures too, not just the perfect demos. Here's the GitHub: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

59,835 views • 8 months ago

JUST IN: Dyna Robotics just published one of the most important research papers in robotics this year. It could fundamentally change how robot foundation models are trained. A scaling law that transfers from human video to robot performance. Dyna-2 is out and it's 🔥 Here's what that means in plain terms. Dyna-2 was pre-trained on ONE MILLION hours of egocentric human video, 170 years of continuous human experience, cooking, folding, assembling, cleaning. And as that human data scaled, robot performance improved. Predictably. Monotonically. Across 39 tasks on two different robot embodiments the model had never seen. → 1,000 hours pre-training → 20% normalised task performance → 10,000 hours → 28% → 100,000 hours → 45% → 1,000,000 hours → 53% Human video exists at effectively unlimited scale. Every cook, every factory worker, every craftsperson wearing a camera is generating training data for future robots. But the finding that stunned even the researchers, world modeling is what makes the transfer work. A model trained to predict future video AND actions massively outperforms one trained on actions alone. Video is the new scaling axis for robotics. One more jaw-dropping data point. 13 minutes of teleoperation data was enough to fine-tune Dyna-2 to open a bottle cap using two five-fingered robot hands. The robots are coming, and they're learning from us directly :D Read more here: Congrats Jason Ma and team! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →
1:35

JUST IN: Dyna Robotics just published one of the most important research papers in robotics this year. It could fundamentally change how robot foundation models are trained. A scaling law that transfers from human video to robot performance. Dyna-2 is out and it's 🔥 Here's what that means in plain terms. Dyna-2 was pre-trained on ONE MILLION hours of egocentric human video, 170 years of continuous human experience, cooking, folding, assembling, cleaning. And as that human data scaled, robot performance improved. Predictably. Monotonically. Across 39 tasks on two different robot embodiments the model had never seen. → 1,000 hours pre-training → 20% normalised task performance → 10,000 hours → 28% → 100,000 hours → 45% → 1,000,000 hours → 53% Human video exists at effectively unlimited scale. Every cook, every factory worker, every craftsperson wearing a camera is generating training data for future robots. But the finding that stunned even the researchers, world modeling is what makes the transfer work. A model trained to predict future video AND actions massively outperforms one trained on actions alone. Video is the new scaling axis for robotics. One more jaw-dropping data point. 13 minutes of teleoperation data was enough to fine-tune Dyna-2 to open a bottle cap using two five-fingered robot hands. The robots are coming, and they're learning from us directly :D Read more here: Congrats Jason Ma and team! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

23,576 views • 1 month ago

Introducing FLUX-mimic, a next-generation Video-Action Model for general purpose dexterity, developed in partnership with Black Forest Labs. Late last year we published mimic-video and introduced Video-Action Models (VAM): a new family of robotics foundation models built on top of video generation models. We showed that robot control reduces to visual prediction, and that robot capability is downstream of improvements in video modeling accuracy. The obvious implication was that advances in the video modeling frontier would directly translate to increased capabilities in end-to-end robot learning. FLUX-mimic is that thesis at frontier scale: We've applied our VAM architecture to the strongest video backbone available today, FLUX 3 from Black Forest Labs, and trained it on data from our own robots and wearables. General-purpose dexterity, running on a single GPU on premises. Because the model already understands world dynamics, it needs far fewer demonstrations to learn a new task. This is game-changing for our mission to deploy robots to factory floors, where industrial robot data is scarce and expensive to collect. We're now testing and deploying FLUX-mimic with manufacturing leaders like Audi USA, on complex, multi-step manipulation long considered impossible for conventional automation.
3:49

Introducing FLUX-mimic, a next-generation Video-Action Model for general purpose dexterity, developed in partnership with Black Forest Labs. Late last year we published mimic-video and introduced Video-Action Models (VAM): a new family of robotics foundation models built on top of video generation models. We showed that robot control reduces to visual prediction, and that robot capability is downstream of improvements in video modeling accuracy. The obvious implication was that advances in the video modeling frontier would directly translate to increased capabilities in end-to-end robot learning. FLUX-mimic is that thesis at frontier scale: We've applied our VAM architecture to the strongest video backbone available today, FLUX 3 from Black Forest Labs, and trained it on data from our own robots and wearables. General-purpose dexterity, running on a single GPU on premises. Because the model already understands world dynamics, it needs far fewer demonstrations to learn a new task. This is game-changing for our mission to deploy robots to factory floors, where industrial robot data is scarce and expensive to collect. We're now testing and deploying FLUX-mimic with manufacturing leaders like Audi USA, on complex, multi-step manipulation long considered impossible for conventional automation.

mimic

118,128 views • 1 month ago

Robotics fundamentally involves understanding the dynamics of how things change in the world in response to action and force. This is impossible to learn from static images; instead, it’s far more effective and more data-efficient to learn from video. Elvis Nava joins us to talk about @mimicrobotic. One of the key findings from mimic-video is that pretraining on webscale video allows robots to learn physics priors; as a result, policies train faster, generalize better, and are capable of more impressive dexterity, versus training on static images or image-language pairs as per a VLM. Watch Episode #81 of RoboPapers with Michael Cho - Rbt/Acc and Chris Paxton to learn more!
1:08:43

Robotics fundamentally involves understanding the dynamics of how things change in the world in response to action and force. This is impossible to learn from static images; instead, it’s far more effective and more data-efficient to learn from video. Elvis Nava joins us to talk about @mimicrobotic. One of the key findings from mimic-video is that pretraining on webscale video allows robots to learn physics priors; as a result, policies train faster, generalize better, and are capable of more impressive dexterity, versus training on static images or image-language pairs as per a VLM. Watch Episode #81 of RoboPapers with Michael Cho - Rbt/Acc and Chris Paxton to learn more!

RoboPapers

48,645 views • 3 months ago

🚨 BREAKING: Microsoft's first robotics foundation model! 🤯 Microsoft just announced Rho-alpha (ρα), their first robotics model derived from the Phi series of vision-language models. Rho-alpha translates natural language commands into control signals for robotic systems performing bimanual manipulation tasks. Commands like "push the green button with the right gripper," "pull out the red wire," "flip the top switch on," or "turn the knob to position 5" get executed directly by dual-arm robots. What makes this different from standard vision-language-action (VLA) models is the additional modalities. Rho-alpha is a VLA+ model that adds tactile sensing to the perceptual mix, with plans to incorporate force feedback. On the learning side, the model is designed to continually improve during deployment by learning from human feedback. The training approach combines trajectories from physical demonstrations and simulated tasks with web-scale visual question answering data. Since teleoperation data is scarce and expensive, Microsoft is using NVIDIA Isaac Sim on Azure to generate physically accurate synthetic datasets via reinforcement learning. These simulated trajectories get combined with commercial and open physical demonstration datasets. The model is currently under evaluation on dual-arm setups and humanoid robots. Microsoft is opening an Early Access Program for organizations interested in evaluating Rho-alpha. Robots that can adapt to dynamic situations and human preferences are more useful in real environments and more trusted by the people operating them. Read more here: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →
0:34

🚨 BREAKING: Microsoft's first robotics foundation model! 🤯 Microsoft just announced Rho-alpha (ρα), their first robotics model derived from the Phi series of vision-language models. Rho-alpha translates natural language commands into control signals for robotic systems performing bimanual manipulation tasks. Commands like "push the green button with the right gripper," "pull out the red wire," "flip the top switch on," or "turn the knob to position 5" get executed directly by dual-arm robots. What makes this different from standard vision-language-action (VLA) models is the additional modalities. Rho-alpha is a VLA+ model that adds tactile sensing to the perceptual mix, with plans to incorporate force feedback. On the learning side, the model is designed to continually improve during deployment by learning from human feedback. The training approach combines trajectories from physical demonstrations and simulated tasks with web-scale visual question answering data. Since teleoperation data is scarce and expensive, Microsoft is using NVIDIA Isaac Sim on Azure to generate physically accurate synthetic datasets via reinforcement learning. These simulated trajectories get combined with commercial and open physical demonstration datasets. The model is currently under evaluation on dual-arm setups and humanoid robots. Microsoft is opening an Early Access Program for organizations interested in evaluating Rho-alpha. Robots that can adapt to dynamic situations and human preferences are more useful in real environments and more trusted by the people operating them. Read more here: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

60,985 views • 7 months ago

Brain-controlled exoskeletons to train humanoid robots! 🧠 Fourier just presented human tele-operators using brain control interfaces and exoskeletal arms to train humanoid robots on home tasks. The brain control interface is the interesting part. Instead of using a controller or joystick to teleoperate, the operator's movements and intentions are captured more naturally through the exoskeleton and BCI. This means the demonstrations are more fluid, more human-like, and better suited for training robots to perform delicate home tasks. Multiple tele-operators are simultaneously generating training data across multiple robots. This is how you build the dataset needed for eventual full autonomy, without waiting years for it to arrive. This might be the bridge between "robots that work in controlled environments" and "robots that work in homes." Not full autonomy right away, but trusted human intelligence operating through a robot body, getting better with every task completed. ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →
0:56

Brain-controlled exoskeletons to train humanoid robots! 🧠 Fourier just presented human tele-operators using brain control interfaces and exoskeletal arms to train humanoid robots on home tasks. The brain control interface is the interesting part. Instead of using a controller or joystick to teleoperate, the operator's movements and intentions are captured more naturally through the exoskeleton and BCI. This means the demonstrations are more fluid, more human-like, and better suited for training robots to perform delicate home tasks. Multiple tele-operators are simultaneously generating training data across multiple robots. This is how you build the dataset needed for eventual full autonomy, without waiting years for it to arrive. This might be the bridge between "robots that work in controlled environments" and "robots that work in homes." Not full autonomy right away, but trusted human intelligence operating through a robot body, getting better with every task completed. ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

20,898 views • 7 months ago

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.
0:48

MotionGPT: Human Motion as a Foreign Language paper page: Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multi-modal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models, motion-language pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this "motion vocabulary", we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based question-and-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.

AK

125,353 views • 3 years ago

Robots don’t learn the physical world from clean datasets. They learn from messy decisions, different movements, and thousands of possible ways a task can go wrong. That’s why I think the data problem in Physical AI is much bigger than simply collecting more examples. A robot can look impressive in a controlled demonstration and still struggle when the environment changes. A different angle. A different object. A slightly different position. A movement that doesn’t go exactly as expected. Humans handle these variations naturally because we have years of physical experience to draw from. Robots don’t. They need enormous amounts of training experience to learn how to perceive situations, choose actions, recover from mistakes, and repeat successful behavior. This is where Axis Robotics is taking an interesting approach. Instead of treating simulation as just a place to test robots, Axis is turning it into a data-generation environment for physical intelligence. More than 200K people are now contributing demonstrations in simulation, creating examples of how tasks can be performed. But the important part isn’t simply the number of contributors. It’s what happens to those demonstrations afterward. The interactions are recorded, validated, and written to Base, creating a verifiable trail around the data being generated. That gives the whole process a different structure: Humans demonstrate → simulations capture → data gets validated → training signals accumulate → robots get more experience. And that feedback loop is what interests me. Because scaling Physical AI may not come from finding one perfect algorithm. It may come from building a system capable of producing millions of useful experiences that algorithms can actually learn from. The robot is only as good as the experience available to train it. And if Axis can keep scaling the quality and volume of that experience, it could become one of the important infrastructure layers behind the next generation of capable robots. The race in Physical AI isn’t only about building smarter models. It’s about giving those models enough real-world experience to become smart in the first place. Axis Robotics.
1:51

Robots don’t learn the physical world from clean datasets. They learn from messy decisions, different movements, and thousands of possible ways a task can go wrong. That’s why I think the data problem in Physical AI is much bigger than simply collecting more examples. A robot can look impressive in a controlled demonstration and still struggle when the environment changes. A different angle. A different object. A slightly different position. A movement that doesn’t go exactly as expected. Humans handle these variations naturally because we have years of physical experience to draw from. Robots don’t. They need enormous amounts of training experience to learn how to perceive situations, choose actions, recover from mistakes, and repeat successful behavior. This is where Axis Robotics is taking an interesting approach. Instead of treating simulation as just a place to test robots, Axis is turning it into a data-generation environment for physical intelligence. More than 200K people are now contributing demonstrations in simulation, creating examples of how tasks can be performed. But the important part isn’t simply the number of contributors. It’s what happens to those demonstrations afterward. The interactions are recorded, validated, and written to Base, creating a verifiable trail around the data being generated. That gives the whole process a different structure: Humans demonstrate → simulations capture → data gets validated → training signals accumulate → robots get more experience. And that feedback loop is what interests me. Because scaling Physical AI may not come from finding one perfect algorithm. It may come from building a system capable of producing millions of useful experiences that algorithms can actually learn from. The robot is only as good as the experience available to train it. And if Axis can keep scaling the quality and volume of that experience, it could become one of the important infrastructure layers behind the next generation of capable robots. The race in Physical AI isn’t only about building smarter models. It’s about giving those models enough real-world experience to become smart in the first place. Axis Robotics.

ABBA

90,854 views • 1 day ago

Breaking news: Cosmos 3 is here. They are attempting to do something completely new 🤯 Why is Physical AI much harder than building a chatbot? Understanding the world is not enough, robots need to predict it and act inside it. That's the idea behind NVIDIA Cosmos 3: → Reasoning model understands what's happening from video, images, text, and actions. → World model generates future states of the environment. → Action model generates the actions needed to achieve a goal. Previous systems often stitched these capabilities together using separate models. Cosmos 3 combines them into a single architecture with two components: ▪️ Reasoner Tower Analyzes observations and builds an understanding of objects, motion, interactions, and physical context. ▪️ Generator Tower Uses that understanding to generate future videos and action sequences that obey physical constraints. So Cosmos 3 moves from: Perception → Model A Prediction → Model B Actions → Model C to: Perception + Prediction + Actions → One unified system. The goal is to make robots, autonomous vehicles, and smart environments better at answering three questions: 1. What is happening? 2. What will happen next? 3. What should I do? That's a big shift from today's AI, which mostly focuses on generating text. And check the benchmarks! Physical AI needs to generate decisions that survive contact with the real world. 🚗🤖
0:33

Breaking news: Cosmos 3 is here. They are attempting to do something completely new 🤯 Why is Physical AI much harder than building a chatbot? Understanding the world is not enough, robots need to predict it and act inside it. That's the idea behind NVIDIA Cosmos 3: → Reasoning model understands what's happening from video, images, text, and actions. → World model generates future states of the environment. → Action model generates the actions needed to achieve a goal. Previous systems often stitched these capabilities together using separate models. Cosmos 3 combines them into a single architecture with two components: ▪️ Reasoner Tower Analyzes observations and builds an understanding of objects, motion, interactions, and physical context. ▪️ Generator Tower Uses that understanding to generate future videos and action sequences that obey physical constraints. So Cosmos 3 moves from: Perception → Model A Prediction → Model B Actions → Model C to: Perception + Prediction + Actions → One unified system. The goal is to make robots, autonomous vehicles, and smart environments better at answering three questions: 1. What is happening? 2. What will happen next? 3. What should I do? That's a big shift from today's AI, which mostly focuses on generating text. And check the benchmarks! Physical AI needs to generate decisions that survive contact with the real world. 🚗🤖

Turing Post

13,836 views • 3 months ago

How well do egocentric human videos scale? Apparently, really well. Dyna Robotics has introduced Dyna-2, a world-action model (WAM) pre-trained on more than 1 million hours of real egocentric human video. It learns by jointly predicting the next video frames and the actions, and that video prediction (world modeling) is what carries learning from humans to robots. • More human video predictably improves action prediction. • Task accuracy improves across embodiments (bimanual arms and a humanoid) with zero robot data in pre-training. • Post-trained on just a few hours of robot data, mean task performance rose 20% to 53% as human pre-training scaled 1k to 1M hours. WAM vs VLA, matched apple-to-apple: the WAM hit 1.55x the VLA's success rate and won 65% of head-to-heads. Deployed zero-shot at unseen customer sites, it passed 87% vs the VLA's 46%.
0:28

How well do egocentric human videos scale? Apparently, really well. Dyna Robotics has introduced Dyna-2, a world-action model (WAM) pre-trained on more than 1 million hours of real egocentric human video. It learns by jointly predicting the next video frames and the actions, and that video prediction (world modeling) is what carries learning from humans to robots. • More human video predictably improves action prediction. • Task accuracy improves across embodiments (bimanual arms and a humanoid) with zero robot data in pre-training. • Post-trained on just a few hours of robot data, mean task performance rose 20% to 53% as human pre-training scaled 1k to 1M hours. WAM vs VLA, matched apple-to-apple: the WAM hit 1.55x the VLA's success rate and won 65% of head-to-heads. Deployed zero-shot at unseen customer sites, it passed 87% vs the VLA's 46%.

The Humanoid Hub

15,837 views • 1 month ago

Excited to announce GR00T N1, the world’s first open foundation model for humanoid robots! We are on a mission to democratize Physical AI. The power of general robot brain, in the palm of your hand - with only 2B parameters, N1 learns from the most diverse physical action dataset ever compiled and punches above its weight: - Real humanoid teleoperation data. - Large-scale simulation data: we are open-sourcing 300K+ trajectories! - Neural trajectories: we apply SOTA video generation models to “hallucinate” new synthetic data that features accurate physics in pixels. Using Jensen’s words, “systematically infinite data”! - Latent actions: we develop novel algorithms to extract action tokens from in-the-wild human videos and neural generated videos. GR00T N1 is a single end-to-end neural net, from photons to actions: - Vision-Language Model (System 2) that interprets the physical world through vision and language instructions, enabling robots to reason about their environment and instructions, and plan the right actions. - Diffusion Transformer (System 1) that “renders” smooth and precise motor actions at 120 Hz, executing the latent plan made by System 2. We deploy N1 on GR1 robot, 1X Neo robot, and a large collection of simulation benchmarks. N1 achieves up to +30% boost in diverse manipulation tasks for household and industrial settings. While humanoid robots are the main focus of N1, our model also supports cross-embodiment. We finetune it to work on the $110 HuggingFace LeRobot SO100 robot arm! Open robot brain runs on open hardware. Sounds just right. Let’s solve robotics, together, one token at a time. Links to our Whitepaper, Github repo, HuggingFace model, and open dataset page in the thread: 🧵
2:21

Excited to announce GR00T N1, the world’s first open foundation model for humanoid robots! We are on a mission to democratize Physical AI. The power of general robot brain, in the palm of your hand - with only 2B parameters, N1 learns from the most diverse physical action dataset ever compiled and punches above its weight: - Real humanoid teleoperation data. - Large-scale simulation data: we are open-sourcing 300K+ trajectories! - Neural trajectories: we apply SOTA video generation models to “hallucinate” new synthetic data that features accurate physics in pixels. Using Jensen’s words, “systematically infinite data”! - Latent actions: we develop novel algorithms to extract action tokens from in-the-wild human videos and neural generated videos. GR00T N1 is a single end-to-end neural net, from photons to actions: - Vision-Language Model (System 2) that interprets the physical world through vision and language instructions, enabling robots to reason about their environment and instructions, and plan the right actions. - Diffusion Transformer (System 1) that “renders” smooth and precise motor actions at 120 Hz, executing the latent plan made by System 2. We deploy N1 on GR1 robot, 1X Neo robot, and a large collection of simulation benchmarks. N1 achieves up to +30% boost in diverse manipulation tasks for household and industrial settings. While humanoid robots are the main focus of N1, our model also supports cross-embodiment. We finetune it to work on the $110 HuggingFace LeRobot SO100 robot arm! Open robot brain runs on open hardware. Sounds just right. Let’s solve robotics, together, one token at a time. Links to our Whitepaper, Github repo, HuggingFace model, and open dataset page in the thread: 🧵

Jim Fan

467,141 views • 1 year ago

Our vision is for AI that uses world models to adapt in new and dynamic environments and efficiently learn new skills. We’re sharing V-JEPA 2, a new world model with state-of-the-art performance in visual understanding and prediction. V-JEPA 2 is a 1.2 billion-parameter model, trained on video, that can enable zero-shot planning in robots—allowing them to plan and execute tasks in unfamiliar environments. Learn more about V-JEPA 2 ➡️ As we continue working toward our goal of achieving advanced machine intelligence (AMI), we’re also releasing three new benchmarks for evaluating how well existing models can reason about the physical world from video. Learn more and download the new benchmarks ➡️
2:03

Our vision is for AI that uses world models to adapt in new and dynamic environments and efficiently learn new skills. We’re sharing V-JEPA 2, a new world model with state-of-the-art performance in visual understanding and prediction. V-JEPA 2 is a 1.2 billion-parameter model, trained on video, that can enable zero-shot planning in robots—allowing them to plan and execute tasks in unfamiliar environments. Learn more about V-JEPA 2 ➡️ As we continue working toward our goal of achieving advanced machine intelligence (AMI), we’re also releasing three new benchmarks for evaluating how well existing models can reason about the physical world from video. Learn more and download the new benchmarks ➡️

AI at Meta

310,549 views • 1 year ago

Jason Ma reveals the first true scaling law in robotics: robots trained on 1 million hours of first-person human video predictably get better without ever seeing robot data "Yesterday we announced our new flagship robot foundation model called Dyna-2. It's very significant for the entire field for many different reasons. First of all, it's the first robot foundation model trained on at least one million hours of data. That in itself is a very challenging infrastructure challenge." "What we demonstrated is that even if the one million hours of data is entirely first-person video of humans doing manipulation tasks, we actually see scaling law for transferring to robot embodiments that the model has never seen before." "By training on more human data, we actually see predictable performance improvement on robots. That's a big deal because the biggest challenge facing robot foundation models is that we don't have enough data." "Unlike language models, there are just not enough robotics data out there, and it's very difficult to collect these robot data. Only by passively observing humans doing things can we scale this kind of data, and this is the first time we showed there is a lot of improvement that can transfer." Jason Ma Dyna Robotics
1:14

Jason Ma reveals the first true scaling law in robotics: robots trained on 1 million hours of first-person human video predictably get better without ever seeing robot data "Yesterday we announced our new flagship robot foundation model called Dyna-2. It's very significant for the entire field for many different reasons. First of all, it's the first robot foundation model trained on at least one million hours of data. That in itself is a very challenging infrastructure challenge." "What we demonstrated is that even if the one million hours of data is entirely first-person video of humans doing manipulation tasks, we actually see scaling law for transferring to robot embodiments that the model has never seen before." "By training on more human data, we actually see predictable performance improvement on robots. That's a big deal because the biggest challenge facing robot foundation models is that we don't have enough data." "Unlike language models, there are just not enough robotics data out there, and it's very difficult to collect these robot data. Only by passively observing humans doing things can we scale this kind of data, and this is the first time we showed there is a lot of improvement that can transfer." Jason Ma Dyna Robotics

MTS

17,637 views • 1 month ago

I don't know if you really understand what these people just did... They built an AI that looks at a few casual real-world videos or photos and instantly turns them into a full 3D playground. You can MOVE THE CAMERA HOWEVER YOU WANT, change the lighting, move objects around, and even let physics play out like in the real world. View events from impossible angles because Atlas knows what was there. This is massive! Let's take robotics. Until now training robots meant collecting tons of expensive real-world data. Now a handful of recordings can become thousands of different simulated situations. Robots can practice the same task in new rooms, with different objects and lighting without anyone having to film it all again. For the industry it means we can train robots way faster and cheaper. Home robots, warehouse robots, whatever... they can learn new environments at scale instead of staying stuck in the lab. This is one of the first real bridges between “cool AI video models” and actual physical machines that have to work in the real world. Blog: The team behind World Labs: Fei-Fei Li, Justin Johnson, Ben Mildenhall and Yunzhu Li. The video is from David Pantera. Follow for more insights into robotics and AI. —— Weekly robotics and AI insights. Subscribe free:
0:49

I don't know if you really understand what these people just did... They built an AI that looks at a few casual real-world videos or photos and instantly turns them into a full 3D playground. You can MOVE THE CAMERA HOWEVER YOU WANT, change the lighting, move objects around, and even let physics play out like in the real world. View events from impossible angles because Atlas knows what was there. This is massive! Let's take robotics. Until now training robots meant collecting tons of expensive real-world data. Now a handful of recordings can become thousands of different simulated situations. Robots can practice the same task in new rooms, with different objects and lighting without anyone having to film it all again. For the industry it means we can train robots way faster and cheaper. Home robots, warehouse robots, whatever... they can learn new environments at scale instead of staying stuck in the lab. This is one of the first real bridges between “cool AI video models” and actual physical machines that have to work in the real world. Blog: The team behind World Labs: Fei-Fei Li, Justin Johnson, Ben Mildenhall and Yunzhu Li. The video is from David Pantera. Follow for more insights into robotics and AI. —— Weekly robotics and AI insights. Subscribe free:

Ilir Aliu

271,773 views • 11 days ago

Most video-action robot models are a content-creation video generator with an action module attached. LingBot-VA 2.0 from Robbyant, a video-action foundation model, throws that starting point out and trains the whole stack natively for control. And it runs closed-loop at a peak 225 Hz. It's so important because A robot cannot move responsively when its controller pauses to imagine the next few frames. LingBot-VA 2.0 predicts during execution, then corrects using each real observation. And it carries only about 13B video parameters while activating roughly 1.9B per token. Bigger robot models usually mean slower reactions, creating a direct conflict between intelligence and control. LingBot-VA 2.0 is trained from scratch for robot control rather than adapted from a video generator built for content creation. Robbyant, an embodied AI company under Ant Group, built it to learn how scenes change under actions, predict what should happen next, and turn those predictions into real-time robot movements. Most video-action systems inherit a tokenizer and video backbone trained mainly to reproduce visual appearance. LingBot-VA 2.0 rebuilds both parts around physical control. Its semantic visual-action tokenizer maps observations toward features from a frozen vision foundation model and learns compact latent actions from frame-to-frame changes using self-supervised inverse and forward dynamics. Unlabeled web video can therefore carry action-relevant training signals without robot action labels. The policy is causal from the start, so every prediction can use only past observations. Its sparse Mixture-of-Experts video backbone has about 13B total parameters, while about 1.9B are active per token, keeping the compute lower during each step. A high-level vision-language planner breaks long tasks into smaller instructions, while the low-level video-action policy handles continuous movement. Foresight Reasoning predicts future visual states while the robot is already acting, then replaces imagined states with every new real observation. Combined with few-step distillation and systems acceleration, the paper reports a peak asynchronous execution frequency of 225 Hz. The model adapts from 10–15 demonstrations, transfers across robot embodiments, and handles some new tasks zero-shot. In the paper’s own evaluations, it reaches 93.6 average on RoboTwin 2.0 and reports stronger real-world results than LingBot-VA and π0.5 across the tested tasks. 🧵 1.
1:17

Most video-action robot models are a content-creation video generator with an action module attached. LingBot-VA 2.0 from Robbyant, a video-action foundation model, throws that starting point out and trains the whole stack natively for control. And it runs closed-loop at a peak 225 Hz. It's so important because A robot cannot move responsively when its controller pauses to imagine the next few frames. LingBot-VA 2.0 predicts during execution, then corrects using each real observation. And it carries only about 13B video parameters while activating roughly 1.9B per token. Bigger robot models usually mean slower reactions, creating a direct conflict between intelligence and control. LingBot-VA 2.0 is trained from scratch for robot control rather than adapted from a video generator built for content creation. Robbyant, an embodied AI company under Ant Group, built it to learn how scenes change under actions, predict what should happen next, and turn those predictions into real-time robot movements. Most video-action systems inherit a tokenizer and video backbone trained mainly to reproduce visual appearance. LingBot-VA 2.0 rebuilds both parts around physical control. Its semantic visual-action tokenizer maps observations toward features from a frozen vision foundation model and learns compact latent actions from frame-to-frame changes using self-supervised inverse and forward dynamics. Unlabeled web video can therefore carry action-relevant training signals without robot action labels. The policy is causal from the start, so every prediction can use only past observations. Its sparse Mixture-of-Experts video backbone has about 13B total parameters, while about 1.9B are active per token, keeping the compute lower during each step. A high-level vision-language planner breaks long tasks into smaller instructions, while the low-level video-action policy handles continuous movement. Foresight Reasoning predicts future visual states while the robot is already acting, then replaces imagined states with every new real observation. Combined with few-step distillation and systems acceleration, the paper reports a peak asynchronous execution frequency of 225 Hz. The model adapts from 10–15 demonstrations, transfers across robot embodiments, and handles some new tasks zero-shot. In the paper’s own evaluations, it reaches 93.6 average on RoboTwin 2.0 and reports stronger real-world results than LingBot-VA and π0.5 across the tested tasks. 🧵 1.

Rohan Paul

11,253 views • 2 months ago

1/7 Physical AI has a data problem. More compute alone won’t teach a robot how to move, grasp, react, or recover from mistakes. LLMs had the entire internet to learn from. Robots need something different: huge amounts of real action data. That’s where Axis Robotics gets interesting. $AXIS is building a data engine for Physical AI, turning robot training into something people can contribute to directly through a browser. The loop is straightforward: Generate → Collect → Process → Deploy → Correct → Repeat. What looks simple on the surface could become something much bigger underneath: a scalable way to produce the training data robots need to become genuinely useful in the physical world. The real unlock for Physical AI might not be another bigger model. It might be better data.
1:10

1/7 Physical AI has a data problem. More compute alone won’t teach a robot how to move, grasp, react, or recover from mistakes. LLMs had the entire internet to learn from. Robots need something different: huge amounts of real action data. That’s where Axis Robotics gets interesting. $AXIS is building a data engine for Physical AI, turning robot training into something people can contribute to directly through a browser. The loop is straightforward: Generate → Collect → Process → Deploy → Correct → Repeat. What looks simple on the surface could become something much bigger underneath: a scalable way to produce the training data robots need to become genuinely useful in the physical world. The real unlock for Physical AI might not be another bigger model. It might be better data.

_Victorad

19,657 views • 18 days ago

I don’t know if we live in a Matrix, but I know for sure that robots will spend most of their lives in simulation. Let machines train machines. I’m excited to introduce DexMimicGen, a massive-scale synthetic data generator that enables a humanoid robot to learn complex skills from only a handful of human demonstrations. Yes, as few as 5! DexMimicGen addresses the biggest pain point in robotics: where do we get data? Unlike with LLMs, where vast amounts of texts are readily available, you cannot simply download motor control signals from the internet. So researchers teleoperate the robots to collect motion data via XR headsets. They have to repeat the same skill over and over and over again, because neural nets are data hungry. This is a very slow and uncomfortable process. At NVIDIA, we believe the majority of high-quality tokens for robot foundation models will come from simulation. What DexMimicGen does is to trade GPU compute time for human time. It takes one motion trajectory from human, and multiplies into 1000s of new trajectories. A robot brain trained on this augmented dataset will generalize far better in the real world. Think of DexMimicGen as a learning signal amplifier. It maps a small dataset to a large (de facto infinite) dataset, using physics simulation in the loop. In this way, we free humans from babysitting the bots all day. The future of robot data is generative. The future of the entire robot learning pipeline will also be generative. 🧵
0:58

I don’t know if we live in a Matrix, but I know for sure that robots will spend most of their lives in simulation. Let machines train machines. I’m excited to introduce DexMimicGen, a massive-scale synthetic data generator that enables a humanoid robot to learn complex skills from only a handful of human demonstrations. Yes, as few as 5! DexMimicGen addresses the biggest pain point in robotics: where do we get data? Unlike with LLMs, where vast amounts of texts are readily available, you cannot simply download motor control signals from the internet. So researchers teleoperate the robots to collect motion data via XR headsets. They have to repeat the same skill over and over and over again, because neural nets are data hungry. This is a very slow and uncomfortable process. At NVIDIA, we believe the majority of high-quality tokens for robot foundation models will come from simulation. What DexMimicGen does is to trade GPU compute time for human time. It takes one motion trajectory from human, and multiplies into 1000s of new trajectories. A robot brain trained on this augmented dataset will generalize far better in the real world. Think of DexMimicGen as a learning signal amplifier. It maps a small dataset to a large (de facto infinite) dataset, using physics simulation in the loop. In this way, we free humans from babysitting the bots all day. The future of robot data is generative. The future of the entire robot learning pipeline will also be generative. 🧵

Jim Fan

165,246 views • 1 year ago

Chinese robotics company Astribot released their latest World-Action Model (WAM), Lumo-2. Technical breakdown: - based on a frozen 🥶 Qwen-3.5 4B VLM - trained in 3 progressive stages: 1. Action is aligned with latent world dynamics (an abstract representation of action). Real-world actions are anchored to physical constraints, while the latent space is guided to focus on motion-relevant changes. This bidirectional relationship makes the model physically grounded -> critical for a world model. 2. Action is aligned with vision and language. Reusing the vision backbone and action encoder from the frozen VLM, the authors add a custom vocabulary (for new actions), a semantic module, an action decoder, and an action projector. This aligns the (new) action representations with the (existing) vision-language semantic space. Most importantly: it builds a direct mapping from natural-language instructions to motor execution. 3. End-to-end training on language, video, and robot data. Only the new modules (everything outside the frozen backbone) are trained end-to-end across temporal reasoning, physical understanding, long-horizon, and dexterous manipulation. At the end of the day, Lumo-2 is not the best on benchmarks, but that's not the point. What's genuinely new: - a way to combine latent world modeling and action generation through progressive alignment - a physically-grounded latent dynamics space - it lifts performance on unseen objects using un-annotated human egocentric video + Vision Pro captures, no special transfer algorithm needed Why it matters: - the whole model is thin trainable adapters (semantic module, action decoder/projector) on a frozen 4B backbone (cheap) - that scale is suited for real-time embedded inference (~2.71× decode speedup, no accuracy loss) - its real moat is long-horizon execution, where the added temporal memory pays off far more than on any other task As a result, this robot can now make your latte (5x sped up video):
0:39

Chinese robotics company Astribot released their latest World-Action Model (WAM), Lumo-2. Technical breakdown: - based on a frozen 🥶 Qwen-3.5 4B VLM - trained in 3 progressive stages: 1. Action is aligned with latent world dynamics (an abstract representation of action). Real-world actions are anchored to physical constraints, while the latent space is guided to focus on motion-relevant changes. This bidirectional relationship makes the model physically grounded -> critical for a world model. 2. Action is aligned with vision and language. Reusing the vision backbone and action encoder from the frozen VLM, the authors add a custom vocabulary (for new actions), a semantic module, an action decoder, and an action projector. This aligns the (new) action representations with the (existing) vision-language semantic space. Most importantly: it builds a direct mapping from natural-language instructions to motor execution. 3. End-to-end training on language, video, and robot data. Only the new modules (everything outside the frozen backbone) are trained end-to-end across temporal reasoning, physical understanding, long-horizon, and dexterous manipulation. At the end of the day, Lumo-2 is not the best on benchmarks, but that's not the point. What's genuinely new: - a way to combine latent world modeling and action generation through progressive alignment - a physically-grounded latent dynamics space - it lifts performance on unseen objects using un-annotated human egocentric video + Vision Pro captures, no special transfer algorithm needed Why it matters: - the whole model is thin trainable adapters (semantic module, action decoder/projector) on a frozen 4B backbone (cheap) - that scale is suited for real-time embedded inference (~2.71× decode speedup, no accuracy loss) - its real moat is long-horizon execution, where the added temporal memory pays off far more than on any other task As a result, this robot can now make your latte (5x sped up video):

Léo

32,331 views • 1 month ago