Archi retweeted
300 hours. 28,993 demonstrations. 59 tasks.
Touch on all 10 fingertips, in every frame.
We're thrilled to share SharpaDex v1.0 with the robotics community! To our knowledge, SharpaDex is the largest publicly available real-robot dexterous-hand manipulation dataset with fingertip tactile sensing, measured by publicly released demonstration hours containing tactile data as of October 2026.
This release brings together:
• Diverse manipulation: tool use, assembly, deformable-object handling, and long-horizon tasks.
• Synchronized sensing: multi-view RGB, fingertip tactile images, fingertip force/torque, and 65-dimensional joint actions.
• Contact-rich data: approximately 195 hours with at least one fingertip in contact, including 55 hours of simultaneous contact across both hands.
• Hierarchical language annotations: task descriptions, 217,642 temporally grounded subtask segments covering 89.3% of frames, and 45 distinct skill labels.
What excites us most is what comes next: researchers using these data to explore how robots can learn from touch, coordinate both hands, and perform increasingly complex tasks.
We hope SharpaDex becomes a useful resource for visual-tactile VLA models, tactile representation learning, and general-purpose dexterous manipulation. We can't wait to see what you build—and learn alongside you.
Available in LeRobot v3.0 and v2.1 formats under CC BY-NC 4.0.
Explore the data, tasks, and annotations through our report and interactive visualizations, or download the dataset to get started:
Dataset on Hugging Face：huggingface.co/datasets/Shar…
Archi retweeted
Check out Haoyang’s blog! I agree with almost every point.
If you’re curious about how Astra-like S2 relates to S1 and S0, and why S0 matters for future humanoids and dexterous hands, this is well worth a read!
My favorite part: “Today's architecture is like one person wearing a blindfold while another watches the environment and tells them how to move their body.”
My blog on System 0 in the Astra era:
As spatial grounding (S1/2) improves, we need dynamics-grounded controllers (S0) and expressive interfaces that let Astra steer physical interaction.
Hope this resonates or sparks disagreement. I’d love your thoughts!
haoyangweng.pages.dev/blog/2…
Archi retweeted
Hey @sundayrobotics, we also tried that! But ACT2 seems to work way better, congrats on the release!
Archi retweeted
Excited to share our recent work, EmbodiedRSI 🤖 — enabling an Embodied Agent Harness to autonomously decide “what to try next and which direction to improve.”
EmbodiedRSI maintains hypotheses for improving code and skills, prioritizes experiments with high information value and low execution cost, and uses real-world execution outcomes to drive the co-evolution of code and skills. Meanwhile, reward-based learning helps identify which experiences are worth retaining and reusing, allowing each interaction to more effectively guide future improvements.
Without updating the underlying robot foundation model, EmbodiedRSI improves the success rate on unseen compositional tasks in RoboCasa365 from 40.1% (best baseline) to 71.3%. When transferred zero-shot to real-world robots, the success rate increases from 46.0% to 71.3%. Under the same budget of 50 interactions, EmbodiedRSI improves the success rate from 27.1% with random exploration to 72.9%.
📄 arXiv: arxiv.org/abs/2610.10498
🌐 Webpage: geeksongs.github.io/agentic_…
Archi retweeted
Today we give Atlas dexterous hands, with the strength to take on real work, and the simplicity for mass manufacturing. Like the body, the hands are designed with high fidelity simulation in mind, to enable sim-to-real RL. If you give dexterous hands to a robot, make them do dexterous things.
More thoughts here
Blog: bostondynamics.com/blog/robo…
Video: youtube.com/watch?v=4whgw2gL…
Archi retweeted
(1/5) Everyone says dexterous manipulation needs real-world data. These policies were trained with none. Our robots mesh gears, thread nuts, and insert pegs zero-shot from RGB, and the same recipe works for locomotion across terrains much harder than before.
The missing piece wasn't a better RL algorithm, but a small tweak to what the robot practices.
Today we're releasing Success-Guided Sampling (SGS), a simple change to the task sampler that unlocks new levels of dexterity and agility with large-scale, end-to-end sim2real RL.
No demonstrations, no per-task reward tuning. Just PPO at scale.
Check out the policies yourself at sgs-rl.github.io! All videos at 1x speed! 🧵 🔈
Archi retweeted
I know teleoperation is old these days but it feels magical when you could teleoperate a humanoid just by wearing a pair of classes (bonus: the robot also sees from glasses too!)
Of course we don't stop here --- we are working on a system to train robots with everyday human experience, without any teleop data. More to come!
Kudos to the team Chuye Zhang @DanielZhenyang @ChuizhengKong
Archi retweeted
Guys! It’s real. It’s real!!
Big Hero 6 is real
and he’s at IROS 2026.
He’s the comfiest most huggable humanoid robot. I want one in my house.
Apparently they’re also going to turn him into a fighting robot!!
Archi retweeted
From printed parts to a walking robot duck. 🦆
Meet XGO-Duck: Arduino UNO Q, 15 servos, and learned motion policies.
Built on Pollen Robotics’ Microduck work.
Explore the hardware, runtime, and training code below. What would you build next?
#maker #3dprinting #robot
Archi retweeted
Happy #IROS2026! Stop by the RAI Institute booth #1102 to grab some stickers and check out some of our custom-built robots including the UMV, Roadrunner, and Koala grippers.
You can also catch Roadrunner in action in the robot demo area. Check the Exhibit Hall demo schedule for more details!
Archi retweeted
Impressive demo from @SharpaRobotics! I like how you can also do the challenge. I can tell you now that the robot is better than me…
#robotics #physicalai
Archi retweeted
WARP is the project I’m most excited about this year: can we learn mobile manipulation policies directly from human data, with zero teleoperation?
The answer is yes. WARP solves a key missing piece: accurate whole-body kinematic retargeting. With it, we can train a wide range of mobile manipulation skills directly from human demonstrations.
There is still work to do, especially on observation-space matching and controller optimization to reach natural human motion and speed. But for the first time, direct human-to-robot policy transfer is starting to look real.
More broadly, this points to a bigger shift: as sensorized human data improves in fidelity, human experience can become a primary source of post-training data for robot policies.
Human data is the cheapest source of whole-body mobile manipulation demos. Turning it into robot data is the hard part.🤖
WARP retargets OFFLINE human motion into replayable robot actions in closed form — precise, consistent, faithful to whole-body intent.
🌍 warp-retargeting.github.io/
Archi retweeted
One lesson we learned from building several mobile manipulation systems:
Coordinate frames aren't just a bookkeeping detail; they fundamentally change how easy a policy is to learn.
e.g., a hand frame makes grasping almost pose-invariant, while a base frame is more natural for motions like "keeping a cup upright while walking."
It's a bit like choosing the right language to describe a motion, and there's no single frame that's best for everything.
So instead of picking one, our Mixture-of-Frames Policy reasons in multiple frames in parallel and combines them during denoising.
Check out @Dian_Wang_ thread for the full story 👇
Long horizon bimanual mobile manipulation requires reasoning in many coordinate frames: base, L/R hands, etc. In which frame would policy work best?
It really depends, so don’t pick. Mixture of Frames Policy: denoise in multiple frame in parallel.
🌐mofpo.github.io (1/9)
Archi retweeted
now opus 5.5 visualized how @physical_int π0.5 inference works on DROID, built from open code and paper
