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。

NEW RESEARCH: You can now create a new robot optimized for any given task! I love this new project by Huy Ha, Shuran Song, and others. Called "Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design", it generates a robot's physical design and its controller together from a task... spec. DEFINITIONS: - Reward function: A scoring rule that assigns a number to how well a behavior achieves the task. Here, it is the objective the generated design is pushed to maximize (e.g., track the target motion with low error). - Tokenizing: dividing continuous or structured data (a robot's links, joints, motor specs, states, actions) into a discrete vocabulary of symbols a transformer can process, the same step that turned pixels and audio into "language" for these models. - Diffusion transformer (DiT): A transformer trained to turn random noise into structured output through iterative denoising. Here, it generates robot bodies and trajectories instead of images. - MuJoCo: The standard fast physics simulator for robotics research (DeepMind-maintained). The Menagerie is its curated zoo of ready-to-use robot models. - CMA-ES: Covariance Matrix Adaptation Evolution Strategy, the workhorse black-box optimizer: it evolves a population of candidate designs, keeps the best, and needs thousands of simulator rollouts. - Bimanual multi-trajectory optimization: Finding one design/controller that performs well across several target motions for a two-armed robot at once, harder than optimizing for a single arm and a single motion. - BERT/MAE masked-modeling trick: Train one model to fill in whatever parts of the input you hide (words for BERT, image patches for MAE); at inference, choosing what to mask chooses the task, so masking the body makes it a designer and masking the actions makes it a controller. In practice, you give it a target end-effector motion and a reward function, and it outputs a complete embodiment (link, joint, motor, and inertial property), as well as a controller to drive it. It works by tokenizing both the body (links/joints/motors) and the dynamics (states/actions) into a compact scheme called RoboTokens, training a diffusion transformer (DiT) over them. The same model predicts dynamics using those predictions ("Dynamics Self-Guidance") to push generated designs toward higher reward at inference time. Masking different token types (using the BERT/MAE masked-modeling trick) lets the one model do three jobs: generate an embodiment, control an arbitrary embodiment, or design one conditioned on a motion. It is trained on 11 robots from the MuJoCo Menagerie (0.65 kg hand to 67.5 kg quadruped, 6–35 joints), and validated in sim and on a physical ALOHA doing cloth flinging. I like the fact that this approach inverts the entire recent robotics ideas: designing a policy for a fixed robot -> designing the robot for a fixed task. Every other approach assumes the body is given and learns a controller. Transformer Transformer takes the task (target motion + reward), then generates the body and controller jointly. In practice, it is a ~180× speedup over the standard optimizer at equal-or-better quality. It reaches "CMA-ES-level quality in seconds" and finishes bimanual multi-trajectory optimization in that is worth underlining nowadays! Also worth mentioning: this is the lab behind UMI and Handroid, that I mentioned here previously! The team seems extremely creative, i love these out-of-the-box approaches. Enjoy watching the demo of robot optimization in 3D, data acquisition, then real-life testing:show more

Léo

1,215 subscribers

24,267 次观看 • 3 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

New model: your robot can now pack your suitcase 🧳 Xiaomi has released a new robot foundation model. Called Xiaomi-Robotics-1, it is designed to have a robot pick things up and move them around. But first, DEFINITIONS: - Mixture-of-Transformers (MoT): An architecture where separate transformer "experts" (e.g., one for vision-language, one for actions) share a single attention stream, so each modality gets specialized parameters without losing joint reasoning. - Vision-language model (VLM): A model that jointly understands images and text. - Diffusion transformer: A transformer trained to turn noise into structured outputs by iterative denoising, here generating robot actions rather than images. - Action chunks: Short sequences of future actions (e.g., the next ~50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: A faster version of diffusion. The model learns a straight-line velocity field from noise to the target action, so it needs only a few integration steps instead of many denoising ones. Its peculiarity comes from its two stage training: 1. 100,000 hours of video shot through a UMI rig: a handheld 3D-printed gripper with a camera, worn by humans doing ordinary tasks in homes, shops, factories and offices. 2. Adapt to actual robot bodies with ~10,000 hours of real-robot data. It replaces the standard approach of teleoperating a real robot for every hour of training data. Its architecture is a Mixture-of-Transformers pairing a pre-trained Qwen3-VL vision-language model with a diffusion transformer that emits action chunks via flow matching, released in 2.6B, 5.1B and 10.5B parameter variants. However, if you read the entire paper ("Scaling VLA Models with over 100K Hours"), you realize that all of the scaling experiments on 20k hours. Therefore the headline "out-of-the-box success climbing 26% → 75% as pre-training data grows" tops out at 100% of 20k hours! What the full corpus does to that curve is never shown -> and this where things would become interesting! Xiaomi's own conclusion is that model size has stopped mattering and data is the binding constraint. The performance gap among different model sizes are less pronounced than those observed across different data scales. This result suggests that model capacity at the billions-parameter scale may already be sufficient to capture the current dataset's distribution. Which further asks the same question: why not use the 100k video hours? Anyway, I would definitely love to have a couple robots at home that can cooperate to pack my suitcase with items relevant to my next destination:

New model: your robot can now pack your suitcase 🧳 Xiaomi has released a new robot foundation model. Called Xiaomi-Robotics-1, it is designed to have a robot pick things up and move them around. But first, DEFINITIONS: - Mixture-of-Transformers (MoT): An architecture where separate transformer "experts" (e.g., one for vision-language, one for actions) share a single attention stream, so each modality gets specialized parameters without losing joint reasoning. - Vision-language model (VLM): A model that jointly understands images and text. - Diffusion transformer: A transformer trained to turn noise into structured outputs by iterative denoising, here generating robot actions rather than images. - Action chunks: Short sequences of future actions (e.g., the next ~50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: A faster version of diffusion. The model learns a straight-line velocity field from noise to the target action, so it needs only a few integration steps instead of many denoising ones. Its peculiarity comes from its two stage training: 1. 100,000 hours of video shot through a UMI rig: a handheld 3D-printed gripper with a camera, worn by humans doing ordinary tasks in homes, shops, factories and offices. 2. Adapt to actual robot bodies with ~10,000 hours of real-robot data. It replaces the standard approach of teleoperating a real robot for every hour of training data. Its architecture is a Mixture-of-Transformers pairing a pre-trained Qwen3-VL vision-language model with a diffusion transformer that emits action chunks via flow matching, released in 2.6B, 5.1B and 10.5B parameter variants. However, if you read the entire paper ("Scaling VLA Models with over 100K Hours"), you realize that all of the scaling experiments on 20k hours. Therefore the headline "out-of-the-box success climbing 26% → 75% as pre-training data grows" tops out at 100% of 20k hours! What the full corpus does to that curve is never shown -> and this where things would become interesting! Xiaomi's own conclusion is that model size has stopped mattering and data is the binding constraint. The performance gap among different model sizes are less pronounced than those observed across different data scales. This result suggests that model capacity at the billions-parameter scale may already be sufficient to capture the current dataset's distribution. Which further asks the same question: why not use the 100k video hours? Anyway, I would definitely love to have a couple robots at home that can cooperate to pack my suitcase with items relevant to my next destination:

Léo

14,771 次观看 • 3 天前

We asked Angus about the future of robotics and AI training. “I built an industrial-grade kinematic solver for 6 degrees of freedom motion, with full path planning and joint control." "You can stream AI model outputs into it, or run it like a traditional industrial robot." "The benefit of that is all of that information you can stream and record from those joints… you can use that as the training data." "Frankly, moving a robot with a lever is really crap data for training robots. But this gives you smooth, high-quality motion, without needing two robots.”

We asked Angus about the future of robotics and AI training. “I built an industrial-grade kinematic solver for 6 degrees of freedom motion, with full path planning and joint control." "You can stream AI model outputs into it, or run it like a traditional industrial robot." "The benefit of that is all of that information you can stream and record from those joints… you can use that as the training data." "Frankly, moving a robot with a lever is really crap data for training robots. But this gives you smooth, high-quality motion, without needing two robots.”

TBPN

18,811 次观看 • 1 年前

Tired of teleoperation? One human video → 1,000s of robot demos. (📍GitHub ) Scaling Robot Data Without Dynamics Simulation or Robot Hardware Real2Render2Real (R2R2R) is a new way to scale robot data without physics simulation or hardware. You take a phone scan + a single monocular human demo. It tracks the motion, renders photorealistic scenes, and generates diverse, robot-agnostic trajectories ready for training. > No teleop, no sim, no robot, just a phone and a video > Train VLA models and diffusion policies directly on the output > Supports multiple robot embodiments with kinematic consistency > 1000s of demos in 1/27 the time of real-world collection Thank you, Max Fu, for sharing!! Project: Paper: Code coming soon: It shows that with the right pipeline, you can scale robot learning data without touching a robot. One of the most interesting directions in scalable robotics today. —— Weekly robotics and AI insights. Subscribe free:

Tired of teleoperation? One human video → 1,000s of robot demos. (📍GitHub ) Scaling Robot Data Without Dynamics Simulation or Robot Hardware Real2Render2Real (R2R2R) is a new way to scale robot data without physics simulation or hardware. You take a phone scan + a single monocular human demo. It tracks the motion, renders photorealistic scenes, and generates diverse, robot-agnostic trajectories ready for training. > No teleop, no sim, no robot, just a phone and a video > Train VLA models and diffusion policies directly on the output > Supports multiple robot embodiments with kinematic consistency > 1000s of demos in 1/27 the time of real-world collection Thank you, Max Fu, for sharing!! Project: Paper: Code coming soon: It shows that with the right pipeline, you can scale robot learning data without touching a robot. One of the most interesting directions in scalable robotics today. —— Weekly robotics and AI insights. Subscribe free:

Ilir Aliu

42,864 次观看 • 6 个月前

Most motion papers tailor one controller to one specific task. This year at SIGGRAPH, our research team asks: can motor control itself be pretrained and reused? Generative Pretrained Controllers, or GPC, turn motor skills into a vocabulary of discrete tokens and train a transformer-based generative controller through next-token prediction. Just like GPT, the same pretrained controller can then be fine-tuned to solve new tasks. Trained on 600+ hours of motion, GPC runs in real-time inside a physics simulation, producing natural and physically grounded behaviors for interactive control.

Most motion papers tailor one controller to one specific task. This year at SIGGRAPH, our research team asks: can motor control itself be pretrained and reused? Generative Pretrained Controllers, or GPC, turn motor skills into a vocabulary of discrete tokens and train a transformer-based generative controller through next-token prediction. Just like GPT, the same pretrained controller can then be fine-tuned to solve new tasks. Trained on 600+ hours of motion, GPC runs in real-time inside a physics simulation, producing natural and physically grounded behaviors for interactive control.

NVIDIA AI

175,407 次观看 • 1 个月前

This is Delta, a deformable multirotor robot developed by scientists at the University of Tokyo that can both fly and roll using its flexible body design. It uses a thrust control system to move efficiently in the air and on the ground, allowing it to adapt to different environments. This capability makes it useful for tasks like disaster response and long-duration operations, introduced in 2024.

This is Delta, a deformable multirotor robot developed by scientists at the University of Tokyo that can both fly and roll using its flexible body design. It uses a thrust control system to move efficiently in the air and on the ground, allowing it to adapt to different environments. This capability makes it useful for tasks like disaster response and long-duration operations, introduced in 2024.

Space and Technology

26,325 次观看 • 3 个月前

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

How do you know your robot actually succeeded? In v0.6.0, you ask a VLM 🤖 LeRobot ships a new unified reward models API with two zero-shot additions. Robometer is a general-purpose reward model trained on 1M+ trajectories across 21 robot embodiments: point it at any LeRobot dataset and it scores task progress and success straight from video and a language instruction, no task-specific training needed. TOPReward goes even further: no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the token "True" given the video and the instruction. Both come with labeling scripts that write per-frame progress curves into your dataset, ready for reward-aware behavior cloning. More on how they work:

LeRobot

22,229 次观看 • 26 天前

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 control in text, it grounds it in video, using large pre-trained video models that already capture physical motion and dynamics. The idea is straightforward: let the video model handle “what will happen next,” and let a smaller control model focus only on turning that visual plan into robot actions. The result is big gains in practice. Robots trained this way need 10× less data, converge twice as fast, and perform better on both simulated benchmarks and real bimanual manipulation tasks. If robots can “imagine” motion using video, control becomes a much simpler problem. Shoutout to Jonas Pai, Liam Achenbach, Oier Mees, Elvis Nava and the rest of the team! Here's the project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

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 control in text, it grounds it in video, using large pre-trained video models that already capture physical motion and dynamics. The idea is straightforward: let the video model handle “what will happen next,” and let a smaller control model focus only on turning that visual plan into robot actions. The result is big gains in practice. Robots trained this way need 10× less data, converge twice as fast, and perform better on both simulated benchmarks and real bimanual manipulation tasks. If robots can “imagine” motion using video, control becomes a much simpler problem. Shoutout to Jonas Pai, Liam Achenbach, Oier Mees, Elvis Nava and the rest of the team! Here's the project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

50,063 次观看 • 7 个月前

New "superman" Four months later, Shanghai-based Matrix Robotics unveiled the Matrix-3--yet another anthropomorphic humanoid robot situated on the left side of the "Uncanny Valley." Unlike Boston Dynamics' Atlas, which is designed for industrial settings, the Matrix-3 is a general-purpose humanoid robot intended initially for service-oriented environments such as supermarkets, hotels, and office buildings. Standing 1.7 meters tall and weighing 65 kilograms, the adult-sized robot features a full-body covering of biomimetic fabric integrated with tactile sensors, as well as humanoid biomimetic muscles. It employs linear joints rather than the mainstream rotary joints. Matrix claims it is capable of pulling or pushing a load of 200 kg. Optimus makes extensive use of linear joints (linear actuators). It is equipped with dexterous hands offering 27 DOF and boasts a battery life of 4 hours. The robot is powered by a model named WAVE(like an E2E World Action Model) Matrix states that it has already established a manufacturing facility in Shanghai's Zhangjiang with an annual production capacity of 10,000 units. The company plans to deliver 1,000 units this year, with the first batch scheduled for delivery by the end of June. Pricing starts at $99,000. Notably, Matrix founder Zhang Haixing previously served as a leader at Tesla China's Design and Research Center, where he was deeply involved in the early-stage R&D and design of the Optimus robot. This company seems to be delivering on the future envisioned for Optimus.

New "superman" Four months later, Shanghai-based Matrix Robotics unveiled the Matrix-3--yet another anthropomorphic humanoid robot situated on the left side of the "Uncanny Valley." Unlike Boston Dynamics' Atlas, which is designed for industrial settings, the Matrix-3 is a general-purpose humanoid robot intended initially for service-oriented environments such as supermarkets, hotels, and office buildings. Standing 1.7 meters tall and weighing 65 kilograms, the adult-sized robot features a full-body covering of biomimetic fabric integrated with tactile sensors, as well as humanoid biomimetic muscles. It employs linear joints rather than the mainstream rotary joints. Matrix claims it is capable of pulling or pushing a load of 200 kg. Optimus makes extensive use of linear joints (linear actuators). It is equipped with dexterous hands offering 27 DOF and boasts a battery life of 4 hours. The robot is powered by a model named WAVE(like an E2E World Action Model) Matrix states that it has already established a manufacturing facility in Shanghai's Zhangjiang with an annual production capacity of 10,000 units. The company plans to deliver 1,000 units this year, with the first batch scheduled for delivery by the end of June. Pricing starts at $99,000. Notably, Matrix founder Zhang Haixing previously served as a leader at Tesla China's Design and Research Center, where he was deeply involved in the early-stage R&D and design of the Optimus robot. This company seems to be delivering on the future envisioned for Optimus.

CyberRobo

40,727 次观看 • 2 个月前

This looks like another step toward removing the data bottleneck in robotics. The bottleneck exists because recording demonstrations is easy, but labeling every grasp, hold, move, and release is slow and expensive. Perceptron Egocentric dropped a new robotics video annotation system. You give it raw robot-camera or first-person/egocentric video. It gives back structured labels that look more like training data: where each small manipulation starts and ends, what the subtask is, what each hand is doing, where the hands are frame by frame, and which hand is left or right. So, Input: raw video of a person or robot doing a task. Output: a machine-readable breakdown of the physical actions inside that video. Basically, it is a machine that converts robot video into policy-training supervision. That means timestamps, subtask boundaries, per-hand actions, and left-right hand grounding. So this will turn large piles of robot or egocentric video into supervision for training robot policies.

This looks like another step toward removing the data bottleneck in robotics. The bottleneck exists because recording demonstrations is easy, but labeling every grasp, hold, move, and release is slow and expensive. Perceptron Egocentric dropped a new robotics video annotation system. You give it raw robot-camera or first-person/egocentric video. It gives back structured labels that look more like training data: where each small manipulation starts and ends, what the subtask is, what each hand is doing, where the hands are frame by frame, and which hand is left or right. So, Input: raw video of a person or robot doing a task. Output: a machine-readable breakdown of the physical actions inside that video. Basically, it is a machine that converts robot video into policy-training supervision. That means timestamps, subtask boundaries, per-hand actions, and left-right hand grounding. So this will turn large piles of robot or egocentric video into supervision for training robot policies.

Rohan Paul

18,903 次观看 • 1 个月前

Let's reverse engineer Disney's adorable, lifelike robot! I couldn't find a whitepaper, but this is how I think it's trained: 1. The emotional behaviors are curated by Disney animation artists, keyframe by keyframe. But it cannot be "rendered" directly on the robot because it doesn't take into account the complex real-world physics. 2. Reinforcement learning (RL) is a great tool for training low-level robot controllers. RL needs a reward function to optimize, and it's typically a task reward (e.g. walk in a straight line as fast as possible). The problem is that RL doesn't know what counts as "natural behavior", and often produces weird-looking body postures that somehow still maximize the reward. This is a human alignment problem just like ChatGPT. 3. Enters Adversarial Motion Prior (AMP): a technique that learns the human preference by training a classifier on what we consider "emotional & cute". In GAN literature, this is called a discriminator. Disney artists are good at creating such a dataset. You can then add AMP as an auxiliary reward in simulation to nudge the robot towards desired behaviors. AMP was developed by Peng et al. 2021 and Escontrela et al. 2022. 4. Add lots of data augmentation to make the controller robust to physical disturbances. In RL, it's called "domain randomization". This is a very powerful technique that bridges the gap between simulator and reality. Previously, OpenAI used domain randomization to train a 5-finger robot hand to manipulate a Rubik's Cube: IEEE news article gave hints about the pipeline: Finally, praying for world peace 🙏. I hope robotics like this will bring more joy to the world.

Let's reverse engineer Disney's adorable, lifelike robot! I couldn't find a whitepaper, but this is how I think it's trained: 1. The emotional behaviors are curated by Disney animation artists, keyframe by keyframe. But it cannot be "rendered" directly on the robot because it doesn't take into account the complex real-world physics. 2. Reinforcement learning (RL) is a great tool for training low-level robot controllers. RL needs a reward function to optimize, and it's typically a task reward (e.g. walk in a straight line as fast as possible). The problem is that RL doesn't know what counts as "natural behavior", and often produces weird-looking body postures that somehow still maximize the reward. This is a human alignment problem just like ChatGPT. 3. Enters Adversarial Motion Prior (AMP): a technique that learns the human preference by training a classifier on what we consider "emotional & cute". In GAN literature, this is called a discriminator. Disney artists are good at creating such a dataset. You can then add AMP as an auxiliary reward in simulation to nudge the robot towards desired behaviors. AMP was developed by Peng et al. 2021 and Escontrela et al. 2022. 4. Add lots of data augmentation to make the controller robust to physical disturbances. In RL, it's called "domain randomization". This is a very powerful technique that bridges the gap between simulator and reality. Previously, OpenAI used domain randomization to train a 5-finger robot hand to manipulate a Rubik's Cube: IEEE news article gave hints about the pipeline: Finally, praying for world peace 🙏. I hope robotics like this will bring more joy to the world.

Jim Fan

314,654 次观看 • 2 年前

The World Model as NEO's Cognitive Core 1X has revealed a major AI development where the NEO humanoid can translate any natural language prompt into robotic action. It demonstrates this capability even for novel tasks, objects, and environments not found in its robot dataset. - the 1X World Model is trained on internet-scale human interaction videos and fine-tuned with robot data to ground its understanding in physics and in NEO's embodiment - from a simple voice or text prompt, the world model generates a visualization of future actions - a built-in inverse dynamics model then translates these into precise motor movements for NEO

The World Model as NEO's Cognitive Core 1X has revealed a major AI development where the NEO humanoid can translate any natural language prompt into robotic action. It demonstrates this capability even for novel tasks, objects, and environments not found in its robot dataset. - the 1X World Model is trained on internet-scale human interaction videos and fine-tuned with robot data to ground its understanding in physics and in NEO's embodiment - from a simple voice or text prompt, the world model generates a visualization of future actions - a built-in inverse dynamics model then translates these into precise motor movements for NEO

The Humanoid Hub

68,453 次观看 • 6 个月前

🇨🇳 Another great Chinese Model, OmniHuman-1.5 from ByteDance Turns 1 image plus a voice track into expressive avatar video by pairing a System 1 and System 2 inspired planner with a Diffusion Transformer, Produces coherent motion for over 1 minute with moving camera and multi character scenes. Most avatar models move to the beat of the audio but miss meaning, so gestures feel generic and emotions feel shallow. The fix here is a Multimodal LLM planner that listens to the speech and drafts a structured plan describing intent, emotions, beats, and high level actions, which gives the motion engine clear semantic targets instead of only rhythm. The motion engine is a Multimodal Diffusion Transformer that fuses the plan with audio, the single reference image, and optional text prompts, then synthesizes continuous body, face, and head motion that matches both words and tone. A key trick is a Pseudo Last Frame, a synthetic target that summarizes the next expected state, which stabilizes fusion across modalities and keeps motion consistent over long spans. From just 1 image and speech, the system outputs speaking avatars with synchronized lips, context aware gestures, and continuous camera movement, and it also supports multi character interactions without manual choreography. Reported results show strong lip sync accuracy, high video quality, natural motion, and close match to text prompts, and the same setup works on nonhuman characters too.

Rohan Paul

63,859 次观看 • 11 个月前

Let's reverse engineer this demo. You need 3 things: (1) robust hardware and motor designs that treat simulation as first-class citizen; (2) a human motion capture ("mocap") dataset, such as those for film and gaming characters; (3) massively parallel RL training in GPU-accelerated simulation. Last October, our team trained a 1.5M parameter foundation model called HOVER for such agile motor control. It follows this recipe, roughly speaking (details in thread): (1) Simulation used to be an after-thought. Now, it has to be part of the hardware design process. If your robot doesn't simulate well, you can kiss RL goodbye. Hardware-simulation co-design is a very interesting emergent topic that only becomes meaningful with today's compute capability. (2) Human mocap dataset to produce natural-looking walking and running gaits. That's one huge advantage of using humanoid robot - you get to imitate from tons of human motions that were originally captured for movies or AAA games. At least 3 ways to use the data: - For initialization: pre-train the neural net to imitate human, and then finetune it into the robot form factor with physics turned on; - For reward function: penalize any deviations from the target pose; - For representation learning: treat the human poses as a "motion prior" to constrain the space of robot behaviors. (3) Shove the above into Isaac sim, add a lot of randomization, pump it through PPO, throw in a bunch of GPUs, and then watch Netflix till loss converges. If you have an urge to comment this is CGI, let me save you a few keystrokes — many academic labs now own the G1 robot in the flesh. See our team's HOVER work in the thread: 🧵

Let's reverse engineer this demo. You need 3 things: (1) robust hardware and motor designs that treat simulation as first-class citizen; (2) a human motion capture ("mocap") dataset, such as those for film and gaming characters; (3) massively parallel RL training in GPU-accelerated simulation. Last October, our team trained a 1.5M parameter foundation model called HOVER for such agile motor control. It follows this recipe, roughly speaking (details in thread): (1) Simulation used to be an after-thought. Now, it has to be part of the hardware design process. If your robot doesn't simulate well, you can kiss RL goodbye. Hardware-simulation co-design is a very interesting emergent topic that only becomes meaningful with today's compute capability. (2) Human mocap dataset to produce natural-looking walking and running gaits. That's one huge advantage of using humanoid robot - you get to imitate from tons of human motions that were originally captured for movies or AAA games. At least 3 ways to use the data: - For initialization: pre-train the neural net to imitate human, and then finetune it into the robot form factor with physics turned on; - For reward function: penalize any deviations from the target pose; - For representation learning: treat the human poses as a "motion prior" to constrain the space of robot behaviors. (3) Shove the above into Isaac sim, add a lot of randomization, pump it through PPO, throw in a bunch of GPUs, and then watch Netflix till loss converges. If you have an urge to comment this is CGI, let me save you a few keystrokes — many academic labs now own the G1 robot in the flesh. See our team's HOVER work in the thread: 🧵

Jim Fan

216,139 次观看 • 1 年前

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.

I am stocked to announce that I won the OpenAI Developers Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to Boris and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.

abdel

17,370 次观看 • 1 个月前

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

Sudo su

39,574 次观看 • 2 个月前

ELON MUSK: "Optimus will be the biggest product ever, but it is a very complex problem to solve. It's one of the hardest things to solve, to make an autonomous humanoid robot that can do tasks that, if you simply ask it to do something or show it a video, it can do the task without any programming. No one's ever achieved this, and there are many challenges in the design of the robot to achieve sufficient dexterity and to be very reliable and have long wear and tear, meaning it needs to be out in the field and not break down. Otherwise, you've got sort of a 70 kilogram robot that just flopped over and you've got to carry it out like a body, you know, it doesn't have wheels. And we don't want Optimus to go haywire. So it's a lot of work to scale, to get the design right and to scale production. This is going to be the hardest product to scale manufacturing that we've ever made at Tesla because everything on the robot is new. And the difficulty of scaling the production ramp is proportionate to the newness of the parts in the robot. With Optimus, there is no supply chain. So we've had to build up a supply chain in its entirety or in-house the production. The Optimus production line that we're building out in Fremont in place of what used to be the Model S & X production, it looks incredible. It's quite stunning to see."

ELON MUSK: "Optimus will be the biggest product ever, but it is a very complex problem to solve. It's one of the hardest things to solve, to make an autonomous humanoid robot that can do tasks that, if you simply ask it to do something or show it a video, it can do the task without any programming. No one's ever achieved this, and there are many challenges in the design of the robot to achieve sufficient dexterity and to be very reliable and have long wear and tear, meaning it needs to be out in the field and not break down. Otherwise, you've got sort of a 70 kilogram robot that just flopped over and you've got to carry it out like a body, you know, it doesn't have wheels. And we don't want Optimus to go haywire. So it's a lot of work to scale, to get the design right and to scale production. This is going to be the hardest product to scale manufacturing that we've ever made at Tesla because everything on the robot is new. And the difficulty of scaling the production ramp is proportionate to the newness of the parts in the robot. With Optimus, there is no supply chain. So we've had to build up a supply chain in its entirety or in-house the production. The Optimus production line that we're building out in Fremont in place of what used to be the Model S & X production, it looks incredible. It's quite stunning to see."

DogeDesigner

307,563 次观看 • 16 天前

NVIDIA just dropped a crazy new research idea. Get ready. The same next-token trick behind ChatGPT got pointed at a ragdoll. Honestly the result is unsettling. No exotic new architecture. Chop human motion into a vocabulary of tiny movement tokens. Train a plain GPT to guess the next one, decode it into muscle commands, let physics run. Then the strange part shows up. Shove the character and it catches its own step. Knock it flat and it plants a hand and gets back up. NOBODY wrote a reward for standing up. Here is why this rattles robotics people. Every humanoid team right now is buried under hand-tuned controllers, one for each little skill. 🔵 the old way: script every skill and its reward by hand 🔵 this way: pretrain ONE body on a mountain of motion and the instincts come free Falling and getting up was never programmed. It emerged. Movement is just a language you can autocomplete.

NVIDIA just dropped a crazy new research idea. Get ready. The same next-token trick behind ChatGPT got pointed at a ragdoll. Honestly the result is unsettling. No exotic new architecture. Chop human motion into a vocabulary of tiny movement tokens. Train a plain GPT to guess the next one, decode it into muscle commands, let physics run. Then the strange part shows up. Shove the character and it catches its own step. Knock it flat and it plants a hand and gets back up. NOBODY wrote a reward for standing up. Here is why this rattles robotics people. Every humanoid team right now is buried under hand-tuned controllers, one for each little skill. 🔵 the old way: script every skill and its reward by hand 🔵 this way: pretrain ONE body on a mountain of motion and the instincts come free Falling and getting up was never programmed. It emerged. Movement is just a language you can autocomplete.

clankr

125,756 次观看 • 1 个月前

The next frontier in protein design will not be defined by structure alone, but by the capacity to engineer motion as a first-class principle of function. This is because dynamics is where the real biology lives. Foundational work by Karplus, Levitt & Warshel made clear that chemistry cannot be understood without motion, mechanism, and scale. Gō, Brooks & others showed that proteins possess characteristic collective motions - low-frequency normal modes that capture how whole molecules bend, breathe, and fluctuate. Frauenfelder then sharpened the picture further: proteins are not static objects occupying a single minimum, but dynamic ensembles traversing rugged energy landscapes. And yet the modern AI revolution in protein science has been, above all, a revolution in structure. In our new paper in Matter, Bo Ni and I ask a different question: not what structure will this sequence adopt? but what sequence will realize a prescribed pattern of motion? VibeGen inverts the conventional design paradigm. Rather than treating dynamics as a consequence to be analyzed after the fact, it makes dynamics the design objective from the outset. Using a language diffusion model with two cooperating agents - a designer that proposes sequences and a predictor that critiques them against the target motion profile - the system converges on de novo proteins with tailored vibrational behavior. One of the most intriguing results is a form of functional degeneracy - distinct sequences and distinct folds can satisfy the same target dynamical specification. For a given functional pattern of motion, evolution may have sampled only a small region of the physically realizable design space. The space of viable molecular mechanics may be far larger than the repertoire biology happened to discover. We have made "vibe" into a cultural metaphor - something intuitive, affective, subjective. But at the molecular scale, vibe is not metaphor: It is physics. For a protein, the vibe is the pattern of motion itself; the fluctuations, resonances, and collective displacements that determine what the molecule can do.

Markus J. Buehler

89,773 次观看 • 3 个月前

This work makes a humanoid robot do simple parkour moves by looking with a depth camera and choosing the right move on the fly. The big deal is that it turns lots of small human moves into long, real-time robot behavior, without hand-coding every transition or retraining for each new course. A humanoid robot is usually good at steady walking, but it often fails when it has to do fast moves like jumping up, vaulting, or rolling, and then keep going to the next obstacle. The hard part is that you cannot easily collect training data for every possible obstacle shape, distance, and mistake, so robots end up learning a few moves that only work in a narrow setup. This work starts from short clips of real human parkour moves, like stepping over, vaulting, climbing, and rolling. It uses motion matching, which is basically a smart “pick the next clip that fits best right now” search, to stitch those short clips into a long, smooth plan that looks like a human doing a whole course. Then it trains a controller with reinforcement learning (RL), which means the robot learns by trial and error to copy that plan while staying balanced and not falling. After training separate expert controllers for different moves, it compresses them into 1 controller that uses only onboard depth sensing and a simple “go this fast in this direction” command. In real tests on a Unitree G1 humanoid, it can clear multiple obstacles in a row, adapt when obstacles get moved, and climb a wall up to 1.25m.

This work makes a humanoid robot do simple parkour moves by looking with a depth camera and choosing the right move on the fly. The big deal is that it turns lots of small human moves into long, real-time robot behavior, without hand-coding every transition or retraining for each new course. A humanoid robot is usually good at steady walking, but it often fails when it has to do fast moves like jumping up, vaulting, or rolling, and then keep going to the next obstacle. The hard part is that you cannot easily collect training data for every possible obstacle shape, distance, and mistake, so robots end up learning a few moves that only work in a narrow setup. This work starts from short clips of real human parkour moves, like stepping over, vaulting, climbing, and rolling. It uses motion matching, which is basically a smart “pick the next clip that fits best right now” search, to stitch those short clips into a long, smooth plan that looks like a human doing a whole course. Then it trains a controller with reinforcement learning (RL), which means the robot learns by trial and error to copy that plan while staying balanced and not falling. After training separate expert controllers for different moves, it compresses them into 1 controller that uses only onboard depth sensing and a simple “go this fast in this direction” command. In real tests on a Unitree G1 humanoid, it can clear multiple obstacles in a row, adapt when obstacles get moved, and climb a wall up to 1.25m.

Rohan Paul

37,121 次观看 • 5 个月前

We’re excited to share DiT4DiT, an end-to-end Video-Action Model for robot learning that unifies a video Diffusion Transformer and an action Diffusion Transformer in a single cascaded framework. By leveraging the rich spatiotemporal and physical dynamics learned through video generation, rather than static image-text priors, DiT4DiT achieves state-of-the-art results on LIBERO (98.6%) and RoboCasa GR1 (50.8%) with far less training data, delivering over 10× better sample efficiency and up to 7× faster convergence. Real-world deployment on a humanoid robot further shows robust generalization. We believe this is a step toward making video generation a powerful backbone for robot policy learning. This work builds upon the brilliant foundations laid by Nvidia's GR00T and Cosmos. Project: Paper: Code: Coming soon. In the meantime, you can ask your coding agent to reproduce the method based on GR00T/Cosmos.

We’re excited to share DiT4DiT, an end-to-end Video-Action Model for robot learning that unifies a video Diffusion Transformer and an action Diffusion Transformer in a single cascaded framework. By leveraging the rich spatiotemporal and physical dynamics learned through video generation, rather than static image-text priors, DiT4DiT achieves state-of-the-art results on LIBERO (98.6%) and RoboCasa GR1 (50.8%) with far less training data, delivering over 10× better sample efficiency and up to 7× faster convergence. Real-world deployment on a humanoid robot further shows robust generalization. We believe this is a step toward making video generation a powerful backbone for robot policy learning. This work builds upon the brilliant foundations laid by Nvidia's GR00T and Cosmos. Project: Paper: Code: Coming soon. In the meantime, you can ask your coding agent to reproduce the method based on GR00T/Cosmos.

Shuo Yang

31,872 次观看 • 4 个月前