Stone Tao
@Stone_Tao
PhDing @UCSanDiego @NVIDIA @hillbot_ai on scalable robot learning and embodied AI. Co-founded @LuxAIChallenge to build AI competitions. @NSF GRFP fellow
The world’s first humanoid robot capable of swapping its own battery - Chinese company UBTech has unveiled their next-gen humanoid robot, Walker S2.
The world’s first humanoid robot capable of swapping its own battery - Chinese company UBTech has unveiled their next-gen humanoid robot, Walker S2.
we need to move on to focusing on wall time more often. Fast dev cycles are what drive research and engineering to eventually being used by the broader public. Think of the success of isaac gym, that enabled fast iteration on locomotion and now everyone uses a variant of it. and…
x.com/i/article/1946…
this is a common bar for physical AGI and an incredibly difficult one But I generally hope more people target physical “ASI” since we want robots to be able to do tasks much faster/safer than humans
My bar for AGI is far simpler: an AI cooking a nice dinner at anyone’s house for any cuisine. The Physical Turing Test is very likely harder than the Nobel Prize. Moravec’s paradox will continue to haunt us, looming larger and darker, for the decade to come.
might be the first time I made it personally on the trending developers list on github! One of the contributing factors being the lerobot sim2real code

🔥🚨 Preprint alert: Relative Entropy Pathwise Policy Optimization #REPPO 🚨🔥 What if you could have on-policy training without the instability and parameter tuning that plagues #PPO? What if training with deterministic policy gradient just worked? With our new method it does!
more visual sim2real reproduction! This time with modifications to the original code
XLeRobot Sim2real RL: Only 1hr RL training in pure simulation (Maniskill) with a 4090 GPU and then we got this. Faster and smoother than most of the VLA models I have seen on LeRobot. Codes and tutorials fully open-sourced: github.com/StoneT2000/ler… Many thanks to @Stone_Tao!
It is a major policy failure that the US cannot accommodate top AI conferences due to visa issues.
Probably the first and only public video from Skild AI: seeing Frankas, UR5s, humanoid arms - all bimanual robotiq 2f85 grippers?
At a robotics lab in Pittsburgh, engineers are building adaptable, AI-powered robots that could one day work where it's too dangerous for humans. The research drew a visit from President Trump, who touted U.S. dominance in AI as companies announced $90 billion in new investments.
XLeRobot 0.2.3 (XLeVR) out! xlerobot.readthedocs.io • VR (Quest 3) Whole-body Control • Reads everything: head and hand poses, joysticks, and all the buttons • Minimal dependencies, web-based • Modular, can use it on other robots • Open-source, easy install, play in 20min
very well written post on motion imitation
🧠With the shift in humanoid control from pure RL to learning from demonstrations, we take a step back to unpack the landscape. 🔗breadli428.github.io/post/lfd/ 🚀Excited to share our blog post on Feature-based vs. GAN-based Learning from Demonstrations—when to use which, and why it…
for robotics it’s also way harder in general vs text. You have way more modalities of inputs, and sometimes robotics data labels not only includes language, but an entire trajectory (generated via teleop for example). it’s quite easy to teleop in a way that’s hard to learn from
Only after labelling a dataset by your own you know how dirty it is
I won’t be at #ICML2025 but Adria will! He will be presenting the work he did during his time at my lab tomorrow from 11:00AM to 1:30PM, stop by and check it out!
🤖Introducing DEMO3: our new model-based RL framework for multi-stage robotic manipulation from visual inputs and sparse rewards. 🧵🔽 📜 Paper: [arxiv.org/abs/2503.01837] 🌍 Project Page: [adrialopezescoriza.github.io/demo3/] 💻 Code: [github.com/adrialopezesco…]
simulation is the way, thank you for having me on!
🎙️The full Robot Mind episode with Stone Tao is out now. A thoughtful look at what it really takes to advance embodied AI, from smarter simulation and stronger benchmarks to self-evolving training.
another reproduction! who will be the first to try and solve a harder task 👀
100% sim2real, no teleoperation data collected had trouble picking up the block until i changed the brick color to be brighter and more similar to the colors in simulation. @Stone_Tao 's repo github.com/StoneT2000/ler…
"You gotta be Elon kind of rich to explore how data scaling laws apply to robotics — because real-world data doesn’t scale easily." 🎙️Teaser from our upcoming Robot Mind podcast with @Stone_Tao , 2nd-year PhD at UC San Diego and incoming NVIDIA intern. Stone unpacks why scaling…
My latest post: The American DeepSeek Project Build fully open models in the US in the next two years to enable a flourishing, global scientific AI ecosystem to balance China's surge in open-source and an alternative to building products ontop of leading closed models.
Awesome to see people reproducing the accessible @LeRobotHF zero-shot sim2real project! Trained for just 90 min in ManiSkill and deployed directly in real. Sim2real is not easy, but very rewarding when it works Original post by Jianwei Zhang on LinkedIn linkedin.com/posts/jianwei-…
Code/tutorial released for🤗LeRobot zero-shot visual sim2real Reinforcement Learning! Train your SO100 robot in ManiSkill's fast simulator+renderer💨 and deploy a cube picking model zero-shot. Just in time for the LeRobot Hackathon, happy hacking! github.com/StoneT2000/ler…