Early in-house test. Task 1: Aero Hand Open grabs a handful of screws from one bin and drops them into another. Task 2: with a power screwdriver, it fastens a screw on another hand. 1x speed. Robot, hand, and policy are all ours.
First SLAM navigation test of our wheeled humanoid , running through the office @ChestnutRobotic .
U-turn included, no human in the loop.
Wheels first. Hands next.
That's our Aero Hand Open running in this active UMI rig. we kept it under 400g and packaged all the motors inside the palm so it mounts onto any wrist with no kinematic re-tuning. handheld setups like this were one of the use cases we built around. great to see it in this kind of build. @kukreja_abhinav ping us if you need any further support, dms open.
High-quality dexterous robot data is expensive to scale.
That is the bottleneck we are focused on at Chestnut Robotics.
As our co-founder and CTO @JoeDong233873 puts it, our approach is to convert human dexterous manipulation data into robot-quality data, starting from structured industrial environments where robots can create real value first.
Inspired by Generalist AI’s new piece:
Robot learning should be more goal-driven, not model-driven.
World models and VLAs both matter. But today’s biggest bottleneck is still data — especially for dexterous manipulation.
The winner in robotics may not just have a better model.
It may have a better data engine.
open.substack.com/pub/joedon…
Finger tutting looks easy until you try to engineer it. Every finger moves on its own axis. Every joint hits the beat. The palm holds steady through all of it.
So we built a hand that could.
Built to work. Also built to dance.
Dexterous hands work great in controlled settings. We wanted to know what happens when the environment fights back.
So we let a toddler loose in the living room, then gave the Aero Hand Open one instruction: "put the toys back in the bin."
Fully autonomous from there. VLA model perceives the scene, plans each grasp in real time, adapts finger trajectories to objects it has never trained on. Stuffed animals, toy trucks, rattles, all soft and irregular, scattered randomly across a play mat.
No fixtures. No staging. Just a real room, real mess, real manipulation.
Fully open source.
#robotics#embodiedAI#opensource
Sunday just raised $165M 🤖
This validates UMI-based in-the-wild data collection as THE winning strategy for robot manipulation.
But picking up objects is just the beginning.
Next frontier: dexterous manipulation
→ In-hand reorientation (rotating a pen, unscrewing a cap)
→ Tool use designed for human ergonomics (scissors, screwdrivers, knives)
→ Contact-rich bimanual coordination (tying knots, opening jars)
The hard part isn't the model architecture — it's capturing rich hand demonstration data at scale.
Whoever builds the best glove multi-fingered hand capture pipeline wins.
Same UMI playbook. Much harder input space. 🦾
#robotics#dexteroushand#humanoid
We have some exciting news to share! 🚀 Today, TetherIA is officially rebranding to Chestnut Robotics. 🌰🤖 As our focus on advanced robotics has grown, we realized we needed an identity that truly reflects our vision for the future—sturdy, grounded, and innovative. Same dedicated team. Same core mission. Brand new look. Welcome to the next chapter. Let us know what you think of our new logo below! 👇 #ChestnutRobotics#Rebrand#Robotics
1,000 units shipped. Zero marketing spend.
Aero Hand Open is proving that in robotics, product speaks louder than ads.
Every hand you see here found its way to a lab, a factory, or a research team — purely by word of mouth.
This is just the beginning.
#DexterousHand#Robotics
100% behind our CEO @Evan_Tao_T.We’re building the full stack to unlock real-world dexterous robotics, with official deployment coming in 2026.Follow us for every step of this journey.
#robotics#AI#buildinginpublic
I spent years at Tesla Optimus and Apple Vision Pro pushing the limits of hardware.
Here's what I learned: the bottleneck for real-world robots isn't just the hardware — it's the data and the model.
We are here to fix it.
This is what we're building: → A world action model for dexterous manipulation → A low-cost, high-quality in-the-wild data collection system
Goal: real-world deployment in 2026.
This is day one.
#robotics#manipulation#buildingInPublic
Respect, Remi! Seeing you push our hand to its limits is exactly why we built it. Let’s skip the small talk and dive into co-creation. What’s your next wild move? We’d love to team up and build something truly insane together. 🦾🔥
🚀 Big news! We just received two Aero Open Hands from @TetherIA_ai ! 🤖👐
Get ready—new hand-based benchmarks are coming soon to RoboChallenge. Can't wait to put these to the test! Stay tuned. 👀
We built a high-fidelity, cable-driven robotic hand in MuJoCo — accurately reproducing the full internal tendon system 🤖🧩
It’s now officially part of the MuJoCo Menagerie.
On top of it, we developed a zero-shot RL training & real-world deployment pipeline based on MuJoCo Playground.
Here’s a demo of Z-axis rotation, powered by tendon actuation ⚙️👇
💻 GitHub: tetheria.github.io/aero-hand…
🔗 Learn more: shop.tetheria.ai/#robotics#AI#reinforcementlearning#mujoco#opensource#sim2real#dexteroushand