What is true on Hands?

Figure held an all-hands this week, and I've decided to share it here publicly Given our rate of progress, our roadmap has completely changed over the last year, and the team has doubled It's important to get everyone aligned on our mission going forward https://t.co/ScTe5iqVr6

This is wild... 🤯 Unitree just unveiled H2-X, and it’s an absolute beast. Spinning kicks, aerial flips, powerful knee and elbow strikes. It jumps from a 4-meter platform onto mats, rolls with the impact, and gets back up. That whole sequence is a serious test of joint strength and whole-body control. The launch video even casts it as a personal bodyguard. It faces off with an intruder and, in a staged rescue, lifts the front of a car while someone pulls a dummy out from underneath. Unitree says the footage is real and not sped up. It also reports roughly eight hours of battery life while standing. https://t.co/bjxFrmcFWA

How do you _guarantee_ your agent's code is correct? You prove it. Introducing Proof Driven Development Kit (PDD Kit) from @MSFTResearch. PDD Kit helps turn your plan into a formal spec, then proves the agent's code meets that spec for every input. No vibe verification. PDD Kit is integrated with @GitHubNext's Chopin, so you can plan as a team. PDD Kit takes an app's login flow from intent to a verified pull request: It discusses your code, infers your intent as it goes, and proposes a formal spec in Verus, a proof language for Rust. Once you approve the spec, the Verus verifier automatically proves the code meets it. Where proof isn't possible, it falls back to empirical verification with tests and computer-use agents (CUAs). It works on Rust today. No new language, no change in how you produce the underlying code, making formal methods more accessible. The goal is a verified software factory, where engineers review specifications instead of code. Come see it in action at GitHub Universe, Oct 28–29. Swing by the Speakeasy and ask for GitHub Next!

The big news last week was of course the sharp drawdown in high yield data center bonds and what that might mean for the broader AI buildout. I have been eyeing the price analog between the AI boom today and the internet boom-turned-bust in 1998-2000 and have been comforted by the fact that earnings are so strong today that equity valuations are falling instead of rising. We can see that below. 🧵(1/3)

Tesla's Q3 earnings deck: "Optimus production began in Fremont on our high-volume lines. The bots coming off this line are the first production-quality bots that we have manufactured. While we will continue to iterate on the hardware and manufacturing process, we have proven that the current manufacturing process for Optimus is suitable for a commercial product."

You can build an 8-servo robotic hand for under €200 in parts. The AmazingHand by Pollen Robotics has two options: DIY: source 8 Feetech SCS0009 servos, print the mechanical structure in 3D and assemble the linkages using the free GitHub files. Seeed Studio kit around $89, or €119.43 including VAT at RobotShop Europe. It includes the structure, 8 servos, a XIAO-compatible control board, power supply, cables and hardware. Both require assembly. The DIY route gives you the parts list and printable design. The kit bundles the components. Which route would you take to build your own robotic hand?

This might be how machines learn to walk blind. No cameras. No LiDAR. The Noetix N2 (1.2 m, 33 kg) climbs 20 cm curbs using feedback from its joints and an IMU that acts like an inner ear. Every 20 ms, a recurrent neural network reads how the body moves and adjusts its next step based on what it feels. The policy was trained in simulation in just 5 hours on a single GPU. On the real machine it handles: • 15 cm steps • 20° ramps • 20 cm curbs • Uneven stone paths The system relies on proprioception, sensing the body's own movement instead of looking at the terrain. What stands out to me is how much information the joints provide when vision is unavailable.

You can build a dexterous robotic hand for just $314. Prefer it ready to use? The fully assembled version costs about $720. Chestnut Robotics' Aero Hand Open weighs 389 g and has 7 active degrees of freedom across 16 joints, 3 of them in the thumb. Here's the trick cables run through the fingers like tendons so one motor bends several joints at once. Everything is open source CAD STEP files, 3D-printable parts, firmware and a MuJoCo simulation model. It's still a research prototype and its long term reliability hasn't been proven yet.

Why does every joint of this $5,000 humanoid hide a cycloidal gearbox? Because 3D-printed plastic can't take concentrated loads. Spreading the force over many contact points is what makes it work. Full breakdown of Berkeley Humanoid Lite https://t.co/iVnL54aVpK https://t.co/F61w2tpGt8

🚀RP1 is officially here. 🌍Putting control of robotics back in developers’ hands. Introducing RP1 (ROBOTO 01) by RoboParty—a high-performance, fully open-source humanoid robotics platform for frontier robotics research and embodied AI model development. 🤖 Open across the full hardware and software stack: • Robot platform: mechanical CAD, bill of materials, and assembly documentation • Actuators: in-house ROMOMO joint designs and related technical resources • Control: electrical systems, firmware, low-level drivers, and motion control implementations • Development: simulation models, SDKs, data processing tools, and training tools • Engineering foundations: technical documentation and research materials Open the black box. Understand the system. Make it your own. A deviation in motion can originate in the model, control policy, actuator response, or mechanical structure. RP1 opens up the design and implementation of these layers so developers can trace problems, modify the system, and validate changes on real hardware. Open-source algorithms for advanced motion control RP1 has demonstrated continuous walking, recovery from disturbances, getting back up after a fall, and human motion tracking. UFO and MimicLite connect motion learning, control policies, and execution on real hardware. The open toolchain connects data collection, simulation training, hardware deployment, and experimental validation—supporting joint iteration across models, control, and hardware. From RPO to RP1, we’re turning one team’s engineering experience into a starting point for many more. Opening the future of embodied AI to everyone ready to explore it. 🔗RP1 on GitHub: https://t.co/kMqFLi3eJ3 #RoboParty #RP1 #OpenSourceRobotics #EmbodiedAI

Four levels of language annotation on one clip: Environment: Cafe Task: Prepare a latte with latte art Subtask: Pick up the coffee sifter Instruction: Right hand picks up the coffee sifter Plus 3D hand pose (<4 mm), body joints (<2 cm) and hand contact heatmaps. https://t.co/3Hf4q7R08R

Exploration is hard for RL when training high-DoF dex hands ✋ X-Reset makes generalist object-centric RL easier by resetting to human hand-object states 🙋‍♀️ We mitigate the exploration problem, and train a single policy to manipulate a variety of objects to any goal pose 🧵 https://t.co/MtP0URF7nu

Speakers from 1X, Dyna Robotics, Skild AI, Google DeepMind, Physical Intelligence and more. Humanoid robots, dexterous hands and more cool robotics demos on the floor. The robotics community in one floor, and four NVIDIA Jetsons to give away. Agenda: https://t.co/Seo5R501s9

Speakers from 1X, Dyna Robotics, Skild AI, Google DeepMind, Physical Intelligence and more. Humanoid robots, dexterous hands and more cool robotics demos on the floor. The robotics community in one floor, and four NVIDIA Jetsons to give away. Agenda: https://t.co/Seo5R501s9

This robot is making wontons. One hand cups the wrapper while the other adds the filling, then both hands fold it closed. It’s running Natura-1, @naturewill’s 50B-parameter mixture-of-experts model. The team says it was pretrained on roughly a million hours of first-person footage, then adapted with robot data. The team reconstructs 3D hand movements from that footage and pairs them with task descriptions. During training, Natura-1 learns to predict future video and actions together, using past video as context. On the robot, it takes in language and vision and controls 20-DoF hands, with a reported 100 ms observation-to-action loop. The hard part here is supporting the thin wrapper while leaving its edges free to fold, then adjusting the grip as its shape changes. The same foundation model also powers demos of sealing shipping boxes, applying phone screen protectors and writing with a brush. These longer workflows, including making wontons, use additional task-specific training. Nature Will separately reports zero-shot tests with instructions and setups absent from its robot post-training data. In the longer interaction video, the presenter moves plates midway through fruit sorting, and the robot adjusts where it puts the fruit. That video also leaves in missed grasps, retries and spoken corrections. You can see where it adapts and where the presenter prompts it to try again. https://t.co/F1JNNGo2tG

What can a robot do without vision? We gave our robot a sense of touch and showed it can recognize and retrieve objects with its cameras blindfolded. Introducing Rhoda’s new tactile sensor—it feels pressure, texture, vibration, slip, and temperature, just like your fingertips.

🚀RP1 is officially here. 🌍Putting control of robotics back in developers’ hands. Introducing RP1 (ROBOTO 01) by RoboParty—a high-performance, fully open-source humanoid robotics platform for frontier robotics research and embodied AI model development. 🤖 Open across the full hardware and software stack: • Robot platform: mechanical CAD, bill of materials, and assembly documentation • Actuators: in-house ROMOMO joint designs and related technical resources • Control: electrical systems, firmware, low-level drivers, and motion control implementations • Development: simulation models, SDKs, data processing tools, and training tools • Engineering foundations: technical documentation and research materials Open the black box. Understand the system. Make it your own. A deviation in motion can originate in the model, control policy, actuator response, or mechanical structure. RP1 opens up the design and implementation of these layers so developers can trace problems, modify the system, and validate changes on real hardware. Open-source algorithms for advanced motion control RP1 has demonstrated continuous walking, recovery from disturbances, getting back up after a fall, and human motion tracking. UFO and MimicLite connect motion learning, control policies, and execution on real hardware. The open toolchain connects data collection, simulation training, hardware deployment, and experimental validation—supporting joint iteration across models, control, and hardware. From RPO to RP1, we’re turning one team’s engineering experience into a starting point for many more. Opening the future of embodied AI to everyone ready to explore it. 🔗RP1 on GitHub: https://t.co/kMqFLi3eJ3 #RoboParty #RP1 #OpenSourceRobotics #EmbodiedAI

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

(2/6) We didn’t expect a 2-finger gripper to be this dexterous. In tight spaces between neighboring parts, AlphaNIST learned several distinct ways to screw on a nut: tapping with the inner faces of its fingertips, turning with their outer faces, and sweeping circular motions around the nut, each suited to how much room it had.

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

[1/7] We trained a world model on 30,000 hours of first-person video. It learned exactly where the hands go. What happens to the things they touch, much less so. Going from 300 to 30,000 hours did not narrow the gap, and a 4× larger model barely did either. https://t.co/2EBqLD1L0c

Robots are getting closer to using their hands like we do. @naturewill is introducing Natura-1, a general foundation model for the physical world built around dexterous control, zero-shot generalization, and long-horizon execution. Here’s what makes this approach interesting. https://t.co/R7JziHJ2yM

Tesla’s patent US 2026/0310299 A1 published October 8, 2026 describes flexible sensor arrays that conform to the curved surfaces of fingers thumb and palm and can serve as tactile skin for a humanoid robot. The sensors use a sandwich structure two layers with conductive patterns separated by a pressure sensitive material. Where the conductive traces overlap, they form independent sensing pixels. Pressure changes the electrical signal capacitance or resistance, allowing the system to detect both the location and the intensity of contact. A key manufacturing route is printing the layers flat then thermoforming the stack so it wraps around curved finger and palm shapes. This approach is designed to make high coverage tactile sensing more scalable. The filing does not mention Optimus by name and does not confirm that these sensors are used on current units. It also contains no performance results. Elon Musk has repeatedly stated that the hand is one of the hardest parts of Optimus. Does this shift the main bottleneck for useful humanoid hands from actuators and control toward scalable tactile skin manufacturing?

if you are building: - actuators - dexterous hands - training data for physical ai - self-driving labs - magnets - the deployment layer of robotics - reindustrialization - space manufacturing - energy and battery innovation - high-fidelity, physics-based world models - BCI for robot control - open-source frontier robot foundation models - anything frontier and sci-fi in physical AI welcome. this account is for you. tell me what you're working on 👇

Tesla’s patent US 2026/0310299 A1 published October 8, 2026 describes flexible sensor arrays that conform to the curved surfaces of fingers thumb and palm and can serve as tactile skin for a humanoid robot. The sensors use a sandwich structure two layers with conductive patterns separated by a pressure sensitive material. Where the conductive traces overlap, they form independent sensing pixels. Pressure changes the electrical signal capacitance or resistance, allowing the system to detect both the location and the intensity of contact. A key manufacturing route is printing the layers flat then thermoforming the stack so it wraps around curved finger and palm shapes. This approach is designed to make high coverage tactile sensing more scalable. The filing does not mention Optimus by name and does not confirm that these sensors are used on current units. It also contains no performance results. Elon Musk has repeatedly stated that the hand is one of the hardest parts of Optimus. Does this shift the main bottleneck for useful humanoid hands from actuators and control toward scalable tactile skin manufacturing?

India is building its own open-source humanoid robot. Meet JX1, developed by Nick the Legend. The project includes: ‣ CAD files and 3D-printable parts ‣ Mechanical designs and STEP/STL files ‣ Bill of materials ‣ Motor control firmware ‣ MuJoCo simulation and walking controller What’s still missing? The physical robot has not been fully assembled and tested for real-world walking, according to the project. You can study the designs and run simulations, but a complete, tested robot ready to walk is not available. Would you build one?

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

Maybe we should have our own "Millennium Problem of Robotics". Robotics academia should be willing to tackle problems that take 5–10 years to solve. Which long-term research problems deserve more attention? My list so far: • Locomanipulation and whole-body coordination • Dexterous, contact-rich manipulation with hands • Force and tactile sensing for perception, learning, and control • Human–robot interaction and collaboration • Evaluation protocols that measure generalization and reliability • Safe, sample-efficient RL in the real world • Long-horizon planning and memory • Lifelong learning and adaptation after deployment • Failure detection, recovery, and knowing when to ask for help • Co-design of robot hardware and learning algorithms What did i miss?

30 minutes. One take. No cuts. We put Natura-1’s zero-shot generalization to the test with complex spoken instructions and unfamiliar tasks and environments. It responds to new requests, adjusts and retries when an action fails, and keeps going autonomously. https://t.co/1B5fL7alKx

[1/7] We trained a world model on 30,000 hours of first-person video. It learned exactly where the hands go. What happens to the things they touch, much less so. Going from 300 to 30,000 hours did not narrow the gap, and a 4× larger model barely did either. https://t.co/2EBqLD1L0c

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

Nature Will @naturewill just released its general foundation model for the physical world, Natura-1, trained on 1 million hours of egocentric human data. Dexterous manipulation is especially challenging. Objects move and rotate, contact points change, and two hands constantly switch roles. The challenge is maintaining control throughout the interaction, not simply grasping an object. (just like the demo) Bringing dexterous control, zero-shot generalization, and long-horizon execution into one foundation model could further expand the capabilities of general-purpose humanoid robots. Worth noting, Natura-1 is another foundation model trained on nearly 1 million hours of human data. Physical AI may be entering the million-hour era.

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.

The first ever Origami Challenge with @SharpaRobotics at @ieeeiros IROS 2026 concluded. We asked teams to fold a paper airplane to test policies in dexterous manipulation. 14 global teams made it to the final round. Here are the results ↓ https://t.co/IepK1hA8cN

Some thoughts on a robotics conference: the research community increasingly is looking at enabling contact-rich manipulation with dexterous hands, and has the tools to do so https://t.co/zZhiEHLX49

Some thoughts on a robotics conference: the research community increasingly is looking at enabling contact-rich manipulation with dexterous hands, and has the tools to do so https://t.co/zZhiEHLX49

You can now use your old Android phone as the brain of a bipedal humanoid robot. UCLA researchers built PhoneBot a 48 cm robot with 13 motorized joints. The phone handles vision and walking control. It can walk track people and stand back up. Hardware costs around $400, excluding the phone.

WUJI at IROS 2026 | Highlights from Pittsburgh Dexterous manipulation with WUJI Hand 2. Hands-on teleoperation with WUJI Glove. From live demos at our booth to discussions at our salon, thank you to the researchers, developers, and industry colleagues who joined us. Watch our IROS 2026 recap below. #WUJI #IROS2026 #DexterousManipulation

WUJI at IROS 2026 | Highlights from Pittsburgh Dexterous manipulation with WUJI Hand 2. Hands-on teleoperation with WUJI Glove. From live demos at our booth to discussions at our salon, thank you to the researchers, developers, and industry colleagues who joined us. Watch our IROS 2026 recap below. #WUJI #IROS2026 #DexterousManipulation

During the @theXtakeover talk, Elon Musk cuts through the hype around flashy humanoid robot demos with a blunt reality check: "Because the AI chip has got to be able to issue the torque commands and close the loop on vision and touch, so this is all very difficult. When you see these robot demo videos, maybe with rare exception, any complicated-looking robot demo, it’s a canned demo. It’s not real. It’s just basically running a routine. It’s not real AI. So real AI is like—you show it something it’s never seen before, and it can still handle that situation."

Three robotic hands in motion - Tesla Optimus - Unitree Dex5 - Clone Robotics https://t.co/XJZjgtQ3kF https://t.co/j6Itqp4A8p

WUJI at IROS 2026 | Highlights from Pittsburgh Dexterous manipulation with WUJI Hand 2. Hands-on teleoperation with WUJI Glove. From live demos at our booth to discussions at our salon, thank you to the researchers, developers, and industry colleagues who joined us. Watch our IROS 2026 recap below. #WUJI #IROS2026 #DexterousManipulation

Ask a robot to grasp a bottle. What happens before its arm moves? Its camera spots the bottle. A VLA model predicts an action. A controller actuates the joints while new images from the camera guide corrections. Even for this simple task, the robot must make continuous adjustments.

A Chinese robotics model just reached first place on RoboDojo with a 32.26% success rate. ROBOTERA's VPP2 predicts future visual states and generates actions for robot manipulation. It scored 39.26 points on RoboDojo-Sim and achieved 58.5% average success across 10 real-robot task categories. The team has released code and model checkpoints. I find the 32.26% result revealing. Even the leading model still fails most attempts across this difficult simulation benchmark.

🤖 For parallel-jaw grippers, high-level plans can already be executed with mature primitives such as grasping. What about dexterous hands? Introducing GOTT (General Object Trajectory Tracking): object trajectories to dexterous execution with a reusable cross-embodiment contact primitive. The same primitive supports diverse real-world tasks including sweeping, pouring, whiteboard erasing, and hammering. With one checkpoint, GOTT reaches 90-96% grasp-and-lift success across multiple objects on three real-world arm-hand platforms. Check videos 👇

Per-joint Gaussian noise is a terrible way to explore a 20+ DoF hand We introduce EigenDEXplore, which explores along coordinated patterns from human hand motion, while keeping full control of every joint Better learning, from tool use 🔨 to reorientation 🔄 to bimanual 👐 https://t.co/M21OpuLppQ