Léo

@LeoKharon • 1,995 subscribers

Robotics research & updates. Co-host @roboticsstack, the weekly pod on what's actually shipping in 🤖

Shorts

NEW WORLD MODEL: Yann LeCun's team is back with an efficient model! This project involves Yann LeCun, Lukas Kuhn, Lucas Maes, Quentin Le Lidec, and Randall Balestriero. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

NEW WORLD MODEL: Yann LeCun's team is back with an efficient model! This project involves Yann LeCun, Lukas Kuhn, Lucas Maes, Quentin Le Lidec, and Randall Balestriero. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

90,459 просмотров

NEW ROBOT: SOLAR PANEL DEPLOYER 🌞 San Francisco Gritt (a.k.a. Gritt AI) is probably one of the hottest robotics startups you have never heard of. Founded in 2023 by two CMU roboticists, it came out of stealth on July 21, 2026 with a $26M Series A ,led by Obvious Ventures . Gritt does not make robots. It bolts an off-the-shelf Kawasaki robotic arm onto heavy equipment for construction, and runs its own AI to unload utility-scale solar panels. They also carry them, and set them onto metal racking with sub-millimeter precision. A human then has to fasten them. It replaces the manual overhead lifting of ~100-lb glass panels on solar farms. Solar is for now their only market, but they will expand to data centers (of course) and other large infrastructures. Solar was chosen first because it's the most factory-like task on an outdoor site. Gritt is making the bet on buying, while most of its competitors are building the robots. Their own capex is therefore near-zero, and moves the scaling constraint to ops crews and software -> it does not make it necessarily easier! All depends where your strengths lie. Gritt has currently two systems deployed in the field, over signed contracts to install ~2.8–3 GW of solar over 18 months. They plan to reach 48 systems within six months -> a ~24× deployment ramp in half a year! Gritt says the first skills took weeks to train but rebar tying took a single day on the same software pipeline after the solar work. That's a ~10–30× drop in per-skill training cost! And goes against the idea that deployment data is near-worthless, since novelty is the scarce input. Also worth mentioning: Gritt has a Chinese competitors, Trinabot. Trinabot is vertically integrated: tied to Trina, a giant Chinese solar-panel maker. Make the panel and the robot that installs it, while Gritt's hardware-agnostic. Interesting to see another example where China integrates manufacturing plus robot, and the US startup goes capex-light software on rented equipment.

NEW ROBOT: SOLAR PANEL DEPLOYER 🌞 San Francisco Gritt (a.k.a. Gritt AI) is probably one of the hottest robotics startups you have never heard of. Founded in 2023 by two CMU roboticists, it came out of stealth on July 21, 2026 with a $26M Series A ,led by Obvious Ventures . Gritt does not make robots. It bolts an off-the-shelf Kawasaki robotic arm onto heavy equipment for construction, and runs its own AI to unload utility-scale solar panels. They also carry them, and set them onto metal racking with sub-millimeter precision. A human then has to fasten them. It replaces the manual overhead lifting of ~100-lb glass panels on solar farms. Solar is for now their only market, but they will expand to data centers (of course) and other large infrastructures. Solar was chosen first because it's the most factory-like task on an outdoor site. Gritt is making the bet on buying, while most of its competitors are building the robots. Their own capex is therefore near-zero, and moves the scaling constraint to ops crews and software -> it does not make it necessarily easier! All depends where your strengths lie. Gritt has currently two systems deployed in the field, over signed contracts to install ~2.8–3 GW of solar over 18 months. They plan to reach 48 systems within six months -> a ~24× deployment ramp in half a year! Gritt says the first skills took weeks to train but rebar tying took a single day on the same software pipeline after the solar work. That's a ~10–30× drop in per-skill training cost! And goes against the idea that deployment data is near-worthless, since novelty is the scarce input. Also worth mentioning: Gritt has a Chinese competitors, Trinabot. Trinabot is vertically integrated: tied to Trina, a giant Chinese solar-panel maker. Make the panel and the robot that installs it, while Gritt's hardware-agnostic. Interesting to see another example where China integrates manufacturing plus robot, and the US startup goes capex-light software on rented equipment.

45,464 просмотров

New robot hands just dropped! But it seems to be missing a couple fingers. Tacta Systems Tacta Systems is a Palo Alto robotics startup that just came out of stealth, and raised with $75M and a dexterity platform, TactaBot, aimed at high-value manufacturing. TactaBot has three parts: - the Tacta Hand -> a human-scale robotic hand with 15 independently actuated joints, proprietary "Fluidic Tendon" actuation, pitched as reliable for millions of factory cycles. - the Tacta Sensor -> a tiny tactile sensor reading force from 250 Pa to 700,000 Pa, sampling at 400 Hz, resolving temperature to 0.1 °C. - Skill Capture -> a data system built around the Tacta Glove. It contains the same sensor embedded in a glove that factory workers wear. Sounds similar to what mimic is doing! Except with 3 fingers instead of 5. Tacta's vision sees 3 fingers as more than enough to complete all tasks, and less complex to simulate than 5. My opinion: Tacta is very strong at semiconductor/MEMS, their moat lies in their ability to manufacture their touch sensor as large as a grain of sand, not the hand or the model. One caveat though: there is no detail nor explanation of how the 5 finger data acquisition glove maps to the 3 finger hand -> I would love to learn more about it! This is where the interesting part of the tech is imho. If this does not work, then nothing does. Still, I have to admit it looks very cool, slick, and minimal, I love it:

New robot hands just dropped! But it seems to be missing a couple fingers. Tacta Systems Tacta Systems is a Palo Alto robotics startup that just came out of stealth, and raised with $75M and a dexterity platform, TactaBot, aimed at high-value manufacturing. TactaBot has three parts: - the Tacta Hand -> a human-scale robotic hand with 15 independently actuated joints, proprietary "Fluidic Tendon" actuation, pitched as reliable for millions of factory cycles. - the Tacta Sensor -> a tiny tactile sensor reading force from 250 Pa to 700,000 Pa, sampling at 400 Hz, resolving temperature to 0.1 °C. - Skill Capture -> a data system built around the Tacta Glove. It contains the same sensor embedded in a glove that factory workers wear. Sounds similar to what mimic is doing! Except with 3 fingers instead of 5. Tacta's vision sees 3 fingers as more than enough to complete all tasks, and less complex to simulate than 5. My opinion: Tacta is very strong at semiconductor/MEMS, their moat lies in their ability to manufacture their touch sensor as large as a grain of sand, not the hand or the model. One caveat though: there is no detail nor explanation of how the 5 finger data acquisition glove maps to the 3 finger hand -> I would love to learn more about it! This is where the interesting part of the tech is imho. If this does not work, then nothing does. Still, I have to admit it looks very cool, slick, and minimal, I love it:

15,474 просмотров

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

NEW SKILL: Underwater drone crab fishing 🦀 I recently went through a Bilibili (kind of Chinese Youtube) rabbit hole, and found the craziest things. For example here, a Chinese vlogger, who produces content about hiking and camping, showing off his crab fishing skills using an underwater drone. Link in comment

NEW SKILL: Underwater drone crab fishing 🦀 I recently went through a Bilibili (kind of Chinese Youtube) rabbit hole, and found the craziest things. For example here, a Chinese vlogger, who produces content about hiking and camping, showing off his crab fishing skills using an underwater drone. Link in comment

109,162 просмотров • 13 дней назад

Some scary stats on China's militaro-industrail complex's capacity: - China casts more metal products than the next nine countries combined and >5× the US. - Its shipbuilding capacity is ~200× the US (a total-capacity figure, not per-ship speed). - It makes ~90% of the world's commercial drones, and controls ~80% of drone components. - Chinese civilian factories could retool within a year to produce ~1 billion weaponized drones annually using under 1% of existing assembly capacity.

Some scary stats on China's militaro-industrail complex's capacity: - China casts more metal products than the next nine countries combined and >5× the US. - Its shipbuilding capacity is ~200× the US (a total-capacity figure, not per-ship speed). - It makes ~90% of the world's commercial drones, and controls ~80% of drone components. - Chinese civilian factories could retool within a year to produce ~1 billion weaponized drones annually using under 1% of existing assembly capacity.

371,473 просмотров • 2 месяцев назад

NEW ROBOT: First transformer humanoid ! HFUN-TECH released a robot that can autonomously reconfigure between a wheel-foot humanoid and a four-legged quadruped, switching modes on its own based on terrain via a unified control framework. You can even carry it as a suitcase, and ride it to move around! It carries a flip camera, a 45-degree tilted interaction screen and a multi-microphone array, and is pitched for home companionship, content creation and outdoor use. It debuted at the 2026 World Artificial Intelligence Conference (WAIC, July 17-20, Shanghai). I also found an alomst identical robot marketed as the "PrimeBOT Qiyuan T1" from a Shanghai startup PrimeBOT, so I cannot really make sense of it 🤔

NEW ROBOT: First transformer humanoid ! HFUN-TECH released a robot that can autonomously reconfigure between a wheel-foot humanoid and a four-legged quadruped, switching modes on its own based on terrain via a unified control framework. You can even carry it as a suitcase, and ride it to move around! It carries a flip camera, a 45-degree tilted interaction screen and a multi-microphone array, and is pitched for home companionship, content creation and outdoor use. It debuted at the 2026 World Artificial Intelligence Conference (WAIC, July 17-20, Shanghai). I also found an alomst identical robot marketed as the "PrimeBOT Qiyuan T1" from a Shanghai startup PrimeBOT, so I cannot really make sense of it 🤔

31,429 просмотров • 21 дней назад

New framework: Kick down your robot, it will get back up every time 🥋 Chinese startup RoboParty is a Beijing startup founded April 2025 by Huang Yi, originally shipping ROBOTO ORIGIN, the world's first full-stack open-source bipedal humanoid. They released UFO: Unsupervised Reinforcement Learning Framework for Humanoid Control. DEFINITIONS -> what differs is where the learning signal comes from: - SUPERVISED: humans supply the right answers (labels), the model imitates them. - UNSUPERVISED: no answer key, the model finds structure in raw data on its own. - REINFORCEMENT LEARNING: no answer key either, the model tries things and a reward scores each attempt. → UNSUPERVISED RL: trial and error where the agent invents its own rewards, instead of engineers hand-writing one per task. REPRESENTATION LEARNING: compress raw states into a useful internal map. TEMPORAL DISTANCE: distance on that map is "how many steps from A to B." CONTRASTIVE: trained by pulling together what's close in time, pushing apart what isn't. -> CONTRASTIVE TEMPORAL-DISTANCE REPRESENTATION LEARNING: the model builds an internal map of body states where distance means how many steps it takes to get from one to another. It is trained by contrast: states that occur close together in a movement get pulled together in the map, randomly paired states get pushed apart. UFO is an open-source training framework that teaches humanoid robots skills, like getting up, walking, goal-reaching, teleoperation, without reference motions -> no motion-capture or human-video demonstrations to imitate. Its core is TeCH, a contrastive temporal-distance representation-learning algorithm: the robot explores, builds pseudo-goals by temporal rolling, and learns goal-conditioned policies from a single unified progress reward. One framework trains five different robots (Unitree G1/H1, RoboParty RP0/RP1, AgiBot X2) with automatic config conversion in ~2–3 hours per robot! The real novelty here "no demonstrations at all". No data-collection arms race,the dominant humanoid-locomotion recipe is tracking: imitate mocap/retargeted-human reference trajectories. The robot self-generates goals from its own exploration and learns from a progress reward, needing zero reference motion data. Everybody else is fighting over data acquisition, while this team just teleports out of the race entirely (inb4 "competition is for losers 💀 ). This strategy reminds me of the DeepSeek playbook applied to robots: open-source the whole stack to become the global default and commoditize everyone else. RoboParty is giving away hardware and now control software (UFO) to be the Android of humanoids. Yet another reason for the US to ban Chinese open models perhaps 🥶 ? What I also really like about this approach is the cross-embodiment infrastructure, one framework trains Unitree G1/H1, RoboParty RP0/RP1, and AgiBot X2 with automatic configuration conversion. Just like Physical Intelligence, RoboParty seems to place itself as a neutral hardware agnostic middle man. Also woth mentioning: their ability ot perform stable skill injection, e.g. adding a cartwheel without forgetting how to walk. A common failure of RL humanoid policies is that teaching a new agile skill destabilizes the existing ones (catastrophic forgetting). UFO claims you can inject rare motions (cartwheel) without collapsing learned behavior. If it holds, that's a significant incremental/continual skill-learning! But again, I have to underline it: no arXiv, no external validation, no success-rate numbers. -> robotics badely needs an independent unbiased evaluator imho. Still, look at that cool demo: robot is getting kicked and pushed around (serious disturbance) during teleoperation (controlled the person at the back wearing the VR headset), and still managed to always get back up. This is some serious demonstration of stability and robustness!

New framework: Kick down your robot, it will get back up every time 🥋 Chinese startup RoboParty is a Beijing startup founded April 2025 by Huang Yi, originally shipping ROBOTO ORIGIN, the world's first full-stack open-source bipedal humanoid. They released UFO: Unsupervised Reinforcement Learning Framework for Humanoid Control. DEFINITIONS -> what differs is where the learning signal comes from: - SUPERVISED: humans supply the right answers (labels), the model imitates them. - UNSUPERVISED: no answer key, the model finds structure in raw data on its own. - REINFORCEMENT LEARNING: no answer key either, the model tries things and a reward scores each attempt. → UNSUPERVISED RL: trial and error where the agent invents its own rewards, instead of engineers hand-writing one per task. REPRESENTATION LEARNING: compress raw states into a useful internal map. TEMPORAL DISTANCE: distance on that map is "how many steps from A to B." CONTRASTIVE: trained by pulling together what's close in time, pushing apart what isn't. -> CONTRASTIVE TEMPORAL-DISTANCE REPRESENTATION LEARNING: the model builds an internal map of body states where distance means how many steps it takes to get from one to another. It is trained by contrast: states that occur close together in a movement get pulled together in the map, randomly paired states get pushed apart. UFO is an open-source training framework that teaches humanoid robots skills, like getting up, walking, goal-reaching, teleoperation, without reference motions -> no motion-capture or human-video demonstrations to imitate. Its core is TeCH, a contrastive temporal-distance representation-learning algorithm: the robot explores, builds pseudo-goals by temporal rolling, and learns goal-conditioned policies from a single unified progress reward. One framework trains five different robots (Unitree G1/H1, RoboParty RP0/RP1, AgiBot X2) with automatic config conversion in ~2–3 hours per robot! The real novelty here "no demonstrations at all". No data-collection arms race,the dominant humanoid-locomotion recipe is tracking: imitate mocap/retargeted-human reference trajectories. The robot self-generates goals from its own exploration and learns from a progress reward, needing zero reference motion data. Everybody else is fighting over data acquisition, while this team just teleports out of the race entirely (inb4 "competition is for losers 💀 ). This strategy reminds me of the DeepSeek playbook applied to robots: open-source the whole stack to become the global default and commoditize everyone else. RoboParty is giving away hardware and now control software (UFO) to be the Android of humanoids. Yet another reason for the US to ban Chinese open models perhaps 🥶 ? What I also really like about this approach is the cross-embodiment infrastructure, one framework trains Unitree G1/H1, RoboParty RP0/RP1, and AgiBot X2 with automatic configuration conversion. Just like Physical Intelligence, RoboParty seems to place itself as a neutral hardware agnostic middle man. Also woth mentioning: their ability ot perform stable skill injection, e.g. adding a cartwheel without forgetting how to walk. A common failure of RL humanoid policies is that teaching a new agile skill destabilizes the existing ones (catastrophic forgetting). UFO claims you can inject rare motions (cartwheel) without collapsing learned behavior. If it holds, that's a significant incremental/continual skill-learning! But again, I have to underline it: no arXiv, no external validation, no success-rate numbers. -> robotics badely needs an independent unbiased evaluator imho. Still, look at that cool demo: robot is getting kicked and pushed around (serious disturbance) during teleoperation (controlled the person at the back wearing the VR headset), and still managed to always get back up. This is some serious demonstration of stability and robustness!

36,112 просмотров • 1 месяц назад

NEW ROBOT BENCHMARK: If your robot can do Origami, it can do anything! Called The Robotic Origami Challenge, it is a dexterous-manipulation competition and benchmark held at IROS 2026, organized by 13 co-organizers with the Nippon Origami Association as judge and task curator -> pretty cool to have them on board imho. The evaluation consists of single task: a traditional Japanese paper airplane, in exactly six folds, from a 15×15 cm sheet of ≥60 gsm paper, on a competition-supplied standardized rig (bimanual arms + Sharpa Hands), both remotely and on-site. Teams bring policies, not hardware. An "Origami Grand Master" declares pass/fail on crease accuracy, structural fidelity, symmetry and paper integrity. Among passes, faster folds rank higher, with a 10-minute-per-attempt ceiling and flight explicitly irrelevant to the score. -> I find it interesting how they chose to evaluate the task. Quality is a binary pass or fail, therefore speed becomes the only thing graded here. Speed is currently the bottleneck in dexterous manipulation though, so this choice makes sense. I wonder whether there could be finer ways to grade the qaulity of such a creative task though. When it comes to data, registered teams get 500+ teleoperation episodes (six camera streams, 65-D joint state/action, 10-fingertip 6-axis tactile), an NVIDIA Isaac Sim environment with thin-shell paper physics (plastic creasing + fold memory), digital twins of every partner hand, and a remote eval lab (upload a policy, queue an eval, get scored). Still, I think it is a great dexterity benchmark the field badly needs, it supplies the hardware, an outside human judges, and the pass criterion is externally defined -> all three degrees of freedom are checked! Neutral measurement layer, here we go! The task is engineered to be un-gameable and to isolate pure dexterity. A known figure, exactly six folds, judged on creases, with flight explicitly declared irrelevant (the latter makes sense to me). Therefore, this underlines the goal to focus on dexterity, not task-selection or other strategies. I really like origami as an ideal controlled dexterity task: deterministic goal, deformable medium, sequential, bimanual, precision-bound. I am just not quite satisfied again by the binary pass or fail, I think quality of execution could be finely graded! But again, I understand this is not the goal yet. Also interesting to see the Sharpa Hands as de facto standard for everyone. Total land-grab that anoints Sharpa as the reference dexterous hand, also featured in Google's Gemini Robotics 2. By providing the hardware, the benchmark measures software while quietly making "good on Sharpa" the definition of good, and Sharpa gets real world data and feedback as a bonus. That's smart, the data flywheel starts spinning. The provided dataset is the richest tactile-manipulation corpus I have seen yet: 10-fingertip 6-axis tactile, plus plastic creasing and fold memory. 500+ teleop episodes with six camera streams, 65-D joint state/action, and ten fingertip 6-axis tactile sensors. The force/tactile channel are parts of the the benchmark's data, this is the first time I see this. Credits where it's due: organizers include Yang Gao, Noriaki Hirose, Steve Xie, Chris Paxton, Jiafei Duan, Michael Cho - Rbt/Acc, Michael Yuan, Haoquan Fang, and others.

NEW ROBOT BENCHMARK: If your robot can do Origami, it can do anything! Called The Robotic Origami Challenge, it is a dexterous-manipulation competition and benchmark held at IROS 2026, organized by 13 co-organizers with the Nippon Origami Association as judge and task curator -> pretty cool to have them on board imho. The evaluation consists of single task: a traditional Japanese paper airplane, in exactly six folds, from a 15×15 cm sheet of ≥60 gsm paper, on a competition-supplied standardized rig (bimanual arms + Sharpa Hands), both remotely and on-site. Teams bring policies, not hardware. An "Origami Grand Master" declares pass/fail on crease accuracy, structural fidelity, symmetry and paper integrity. Among passes, faster folds rank higher, with a 10-minute-per-attempt ceiling and flight explicitly irrelevant to the score. -> I find it interesting how they chose to evaluate the task. Quality is a binary pass or fail, therefore speed becomes the only thing graded here. Speed is currently the bottleneck in dexterous manipulation though, so this choice makes sense. I wonder whether there could be finer ways to grade the qaulity of such a creative task though. When it comes to data, registered teams get 500+ teleoperation episodes (six camera streams, 65-D joint state/action, 10-fingertip 6-axis tactile), an NVIDIA Isaac Sim environment with thin-shell paper physics (plastic creasing + fold memory), digital twins of every partner hand, and a remote eval lab (upload a policy, queue an eval, get scored). Still, I think it is a great dexterity benchmark the field badly needs, it supplies the hardware, an outside human judges, and the pass criterion is externally defined -> all three degrees of freedom are checked! Neutral measurement layer, here we go! The task is engineered to be un-gameable and to isolate pure dexterity. A known figure, exactly six folds, judged on creases, with flight explicitly declared irrelevant (the latter makes sense to me). Therefore, this underlines the goal to focus on dexterity, not task-selection or other strategies. I really like origami as an ideal controlled dexterity task: deterministic goal, deformable medium, sequential, bimanual, precision-bound. I am just not quite satisfied again by the binary pass or fail, I think quality of execution could be finely graded! But again, I understand this is not the goal yet. Also interesting to see the Sharpa Hands as de facto standard for everyone. Total land-grab that anoints Sharpa as the reference dexterous hand, also featured in Google's Gemini Robotics 2. By providing the hardware, the benchmark measures software while quietly making "good on Sharpa" the definition of good, and Sharpa gets real world data and feedback as a bonus. That's smart, the data flywheel starts spinning. The provided dataset is the richest tactile-manipulation corpus I have seen yet: 10-fingertip 6-axis tactile, plus plastic creasing and fold memory. 500+ teleop episodes with six camera streams, 65-D joint state/action, and ten fingertip 6-axis tactile sensors. The force/tactile channel are parts of the the benchmark's data, this is the first time I see this. Credits where it's due: organizers include Yang Gao, Noriaki Hirose, Steve Xie, Chris Paxton, Jiafei Duan, Michael Cho - Rbt/Acc, Michael Yuan, Haoquan Fang, and others.

28,718 просмотров • 1 месяц назад

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):

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):

32,331 просмотров • 1 месяц назад

Every humanoid joint you've seen is three motors stacked on top of each other, each one carrying the weight of the last. The Robotics Stack sat down with Bryce Cianciotto Bryce Cianciotto, who built the Orbit Actuator that does all three axes with the motors at the base, and the wires routed through the middle. New episode 👇

Every humanoid joint you've seen is three motors stacked on top of each other, each one carrying the weight of the last. The Robotics Stack sat down with Bryce Cianciotto Bryce Cianciotto, who built the Orbit Actuator that does all three axes with the motors at the base, and the wires routed through the middle. New episode 👇

20,269 просмотров • 1 месяц назад

NEW DRONE: the most efficient drone ever made! Luke Maximo Bell, a content creator from South Africa, documents how his father-son team set the Guinness record for the longest flight by an electric multirotor! 4 hours, 21 minutes, 39 seconds on a single charge, it beat the previous official record of 3 hours 12 minutes. Tech specs. The aircraft is a lightweight 3D-printed quadcopter that crawls forward at ~5.5 m/s (12 mph), drawing only ~450–500 W, carried by a single ~5 kg SMC semi-solid-state battery pack rated at 380 Wh/kg. Flown on an open-source ArduPilot / CubePilot Cube stack with RTK GPS for centimetre station-keeping. It reused the T-Motor MN10590 KV motors and 40-inch T-Motor G40 props from its predecessor unchanged. I find it interesting that the same two-person team holds both the fastest and the longest-flying electric multirotor records, opposite extremes of the same physics. Their Peregreen 4 hit 657 km/h (408 mph) in January 2026, a record they took back from a rival and one that surpasses Rolls-Royce's fastest-electric-aircraft mark, now the endurance crown at 4h21m. The load-bearing component is an SMC semi-solid-state pack at 380 Wh/kg, up from the Tattu semi-solid-state NMC packs (~320 Wh/kg) that ran two-in-parallel on version 1. Version 2 flies a single 87 Ah, ~5 kg SMC pack. The SMC pack is heavier, but its higher energy density makes the optimal all-up weight land around 5 kg anyway, and the flight-time jump tracks the energy-density jump worth it. Also interesting imho: the record's binding constraint turned out to be vibration, not battery or aerodynamics. The 40-inch props shook the airframe hard enough that the "vibes" clipped the IMU on the flight controller, breaking autonomous flight entirely (constant starts, stops and errors). It was fixed only by swapping to a Cube Orange Plus flight controller with a dampened, vibration-isolated IMU stack (mounted on 3D-printed TPU) plus a front spar to stiffen the arms. Same motors, same props, same airframe as version 1, so the entire gain came from the battery, the vibration fix and tuning.

NEW DRONE: the most efficient drone ever made! Luke Maximo Bell, a content creator from South Africa, documents how his father-son team set the Guinness record for the longest flight by an electric multirotor! 4 hours, 21 minutes, 39 seconds on a single charge, it beat the previous official record of 3 hours 12 minutes. Tech specs. The aircraft is a lightweight 3D-printed quadcopter that crawls forward at ~5.5 m/s (12 mph), drawing only ~450–500 W, carried by a single ~5 kg SMC semi-solid-state battery pack rated at 380 Wh/kg. Flown on an open-source ArduPilot / CubePilot Cube stack with RTK GPS for centimetre station-keeping. It reused the T-Motor MN10590 KV motors and 40-inch T-Motor G40 props from its predecessor unchanged. I find it interesting that the same two-person team holds both the fastest and the longest-flying electric multirotor records, opposite extremes of the same physics. Their Peregreen 4 hit 657 km/h (408 mph) in January 2026, a record they took back from a rival and one that surpasses Rolls-Royce's fastest-electric-aircraft mark, now the endurance crown at 4h21m. The load-bearing component is an SMC semi-solid-state pack at 380 Wh/kg, up from the Tattu semi-solid-state NMC packs (~320 Wh/kg) that ran two-in-parallel on version 1. Version 2 flies a single 87 Ah, ~5 kg SMC pack. The SMC pack is heavier, but its higher energy density makes the optimal all-up weight land around 5 kg anyway, and the flight-time jump tracks the energy-density jump worth it. Also interesting imho: the record's binding constraint turned out to be vibration, not battery or aerodynamics. The 40-inch props shook the airframe hard enough that the "vibes" clipped the IMU on the flight controller, breaking autonomous flight entirely (constant starts, stops and errors). It was fixed only by swapping to a Cube Orange Plus flight controller with a dampened, vibration-isolated IMU stack (mounted on 3D-printed TPU) plus a front spar to stiffen the arms. Same motors, same props, same airframe as version 1, so the entire gain came from the battery, the vibration fix and tuning.

12,227 просмотров • 1 месяц назад

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:

15,723 просмотров • 1 месяц назад

New research: a hand is now a standalone robot! I love this project involving Sikai Li, Haochen Shi, Shuran Song, Mingyu Ding, and others. Called Handroid, it is an open-source, desktop-scale research robot (0.33 m tall, 2.05 kg, 27 DoF) that is both a dexterous anthropomorphic hand and a walking mini-humanoid. How cool is that ?! The same articulated electromechanical modules reconfigured allow the fingers to literally double as limbs. It is able to perform dexterous grasping, in-hand cube reorientation, pick-and-place, pouring, and bipedal squat/walk/turn, plus long-horizon tasks where it reconfigures its own embodiment (between bipedal humanoid and hand) mid-run (dock, locomote, manipulate). The project publishes GitHub code, OnShape CAD, and a Google-Sheets bill of materials, fully open, 3D printable -> you can build your own todat I really like that it is the most genuinely novel mechanical idea I have seen in a long time: not two robots in a box, but a robot whose fingers become its legs. Worth mentioning: Shuran Song's lab (Shuran Song) created UMI, the cheap hand gripper that became a standard for data acquisition. -> the people who built the dominant data-collection interface are now building cheap reconfigurable hardware. I believe this is a tell about where the frontier academic labs are now pushing the frontier: not another VLA, but better/cheaper embodiments to study on. Handroid's ability to transform between hand-mode and humanoid-mode is a genuinely different capability from every other robot we have seen lately, as most other robots have fixed embodiement, besides the ability to pick and use a tool. Morphology as a controllable variable is genuinely new, afaik. Enjoy watching this little robot as much as I do. Here it is able to plan a long horizon task, decompose it in steps, and behaves sequentially as either a hand or a humanoid (4x sped up):

New research: a hand is now a standalone robot! I love this project involving Sikai Li, Haochen Shi, Shuran Song, Mingyu Ding, and others. Called Handroid, it is an open-source, desktop-scale research robot (0.33 m tall, 2.05 kg, 27 DoF) that is both a dexterous anthropomorphic hand and a walking mini-humanoid. How cool is that ?! The same articulated electromechanical modules reconfigured allow the fingers to literally double as limbs. It is able to perform dexterous grasping, in-hand cube reorientation, pick-and-place, pouring, and bipedal squat/walk/turn, plus long-horizon tasks where it reconfigures its own embodiment (between bipedal humanoid and hand) mid-run (dock, locomote, manipulate). The project publishes GitHub code, OnShape CAD, and a Google-Sheets bill of materials, fully open, 3D printable -> you can build your own todat I really like that it is the most genuinely novel mechanical idea I have seen in a long time: not two robots in a box, but a robot whose fingers become its legs. Worth mentioning: Shuran Song's lab (Shuran Song) created UMI, the cheap hand gripper that became a standard for data acquisition. -> the people who built the dominant data-collection interface are now building cheap reconfigurable hardware. I believe this is a tell about where the frontier academic labs are now pushing the frontier: not another VLA, but better/cheaper embodiments to study on. Handroid's ability to transform between hand-mode and humanoid-mode is a genuinely different capability from every other robot we have seen lately, as most other robots have fixed embodiement, besides the ability to pick and use a tool. Morphology as a controllable variable is genuinely new, afaik. Enjoy watching this little robot as much as I do. Here it is able to plan a long horizon task, decompose it in steps, and behaves sequentially as either a hand or a humanoid (4x sped up):

15,346 просмотров • 1 месяц назад

Больше нет контента для загрузки