Model-based control promises generalization--just add a model! Yet it's never been that easy. This summer, we wanted to show that we our C3+ controller COULD "Push Anything," just from a visual scan of an object. 33 different objects, including 4 simultaneously, 98% success rate

I joined 1X because Bernt & Eric asked me if I could help them make NEO into something simple enough for anyone on earth to use and give it to the world. It felt like having a couple strangers come up to me and offer me a chance to change the world (well it was). I've never

48Nm?

🧵 1/8 If vision at 1X has thought me one thing; it's that it's the most important sensor modality for humanoid robots Kun Lei, Chris Paxton, CIX 🦾, Thomas Wolf, C Zhang, Matt Freed, Bryson Jones, Ari Wasch, Jannik Grothusen

We scaled up an "alternative" paradigm in RL: *divide and conquer*. Compared to Q-learning (TD learning), divide and conquer can naturally scale to much longer horizons. Blog post: seohong.me/blog/rl-withou… Paper: arxiv.org/abs/2510.22512 ↓

Cursed. Read some biokinetics paper claiming that vertical force after landing a jump is 100x body weight. Looking at the unit, it turns out to be 100x body mass (10g). Asked ChatGPT and Gemini if this could be true. Both models: yes coz force happens with a small delta t.

Awwwwwww 🎃

Coupling of agility and constraints always looks cool!

Right timing for Halloween! * This kind of motions hurt hardware with impacts, recommend testing on soft ground.

Most of the 1X hate/hype misses what’s actually happening. This is not a product release! This is an audacious gambit. The AI is unsolved (by anyone), but 1X has started the shotclock. They are betting that autonomy timelines are short / can be shortened by Neo adopters.

jack-o-lora

Still keeping crawl at half-speed but it works with the costume on!

What if robots could improve themselves by learning from their own failures in the real-world? Introducing 𝗣𝗟𝗗 (𝗣𝗿𝗼𝗯𝗲, 𝗟𝗲𝗮𝗿𝗻, 𝗗𝗶𝘀𝘁𝗶𝗹𝗹) — a recipe that enables Vision-Language-Action (VLA) models to self-improve for high-precision manipulation tasks. PLD

A100 is really a cursed GPU

Well... Torque accuracy by actuation or feedback? What does 95% mean for backdrivability? What is nominal load? Repetition and accuracy (does that mean precision?) of what? AI based adaptation means adaptive control with randomization of offsets instead of hardstop calibration?