Manipulation & RL @Tesla_Optimus | CMU Robotics | Sharing what I learn about robots and ML | Opinions my own
California, USA
Joined April 2025
- Tweets263
- Following54
- Followers406
- Likes2K
Pinned Tweet
I work on humanoid robots, and even I can't keep up with the field anymore.
So I built Observatory — an interactive map of every player in robotics: what they're building, how mature it is, how big they are.
Link in comment. Tell me who's missing, and I'll add 👇
Very Impressive
(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! 🧵 🔈
How OpenAi solved the Millennium Navier Stokes problem
youtu.be/ofdMfljL360?is=or81…
Very impressive @SharpaRobotics
Pretty cool demo. The controller gets privileged information about the ball position through Vicon tracking.
We trained a robot to play football.
How? Self-play for 140 years in a virtual World Cup.
Meet the #Messinator:
Self play is a very powerful technique to discover strategies purely from experience. Popular examples, OpenAI’s DOTA and Hide & Seek projects. Our 2020 project on Emergent Multi-Agent Strategies showed symmetrical agents can learn asymmetrical behavior through competition.
Got to attend 2 of the fights last season. Excited for season 2.
ULTIMATE BOTS SEASON 2 IS HERE. 🤖🏆
4 teams. 6 events. $100K on the line.
Teams will earn points across 3 disciplines:
🥊 Martial arts
💃 Dancing
🏁 Obstacle course
Season 2 hits F1 Austin → TwitchCon → IEEE Humanoids → CES → Big Game Weekend → NVIDIA GTC.
Captains + teams announced soon.
Follow for reveals, ticket drops & live streams.
Unitree’s new dexterous hand. In the real footage at the end, we can see dexterity gaps.
1. Not able to pinch.
2. Scissors are pre-loaded into the hand instead of being able to pick up from a table.
I am sure they’re working on improving and am excited for the next iteration.
Unitree Introducing: Unitree Dex5-S Dexterous Hand 22 Degrees of Freedom 1:1 Real-Hand Size👋
Precision biomimetic dexterous hand, price from $6.5K (Tax and Shipping cost excluded), with all 22 joints supporting smooth backdrivability, and each joint equipped with limit impact torque protection.
Well said. Tesla is a unique place with concentrated talent, resources and motivation.
Don't sleep on Tesla in humanoids just because they've gone quiet.
Tesla bet on solving driving intelligence from vision alone, put a powerful inference computer in every car long before FSD was solved or profitable, and built the biggest robotics data flywheel.
General-purpose humanoid needs the same systems, but on steroids. Debate the AI model architectures and actuator designs all you want. They definitely matter for setting the right direction. But what ultimately decides the outcome is the organizational strength that powers every decision.
The trifecta:
• Engineering: fast iteration, bringing new work in-house, pivoting quickly
• Physical ops: manufacturing, data collection, supply chain, repairs, fleet management
• Inference and training compute infra at massive scale
Few companies can bring all three together at scale as well as Tesla.
4 generations of autonomous vehicles.
In 2019 I was at CMU. 2007 DARPA urban challenge winning car was brought on campus alongside an Argo AI vehicle, which looked more polished.
Fast forward, AVs are actually on road. Cybercab looks more polished and advanced than a Waymo.
Robotaxi 🚀
Cybercabs takes over downtown Austin — Fully Autonomous Fleet Operating
🎥 @AdanGuajardo with @artsimage
Fast video models can lead to world models that are practical for robotics. Of course there are other issues such as visual artifacts and physics inconsistency that still need to be solved.
Super impressive. Truly believe that such one shot learning from a single human video is what’ll unlock mass adoption of learnt robot policies in factories. Many interesting extensions are possible. E.g. in context learning instead of gradient steps.
This is so cute
You can now book two humanoids for our cleaning service in SF. One humanoid for 60 minutes is $30, two are $60.
You can book a cleaning at tau-robotics.com. If you don’t have an invite code yet, join the waitlist. Timelapse at 5× speed (~40 minutes in real time).
Claude is so incredible at so many things but then it said this to me “You’ve got nothing to lose by giving it another hour and looking again — you’re asleep anyway.” How can it think that I can check something while I am asleep?
😅
The AI labs finally made peace. Jensen brought the chips.
created in two prompts, zero editing, using Seedance 2.5 on @plotpartyai