GRASP Laboratory Physical Intelligence 🧵6/ PI0 is actually a FPV player: It mainly relies on wrist camera. In fact, it still works even if the side-view camera is blocked. 3rd camera view contains more variations and change, the neural network may focus more on what the gripper see to execute tasks.

𝗔𝗳𝘁𝗲𝗿 𝟭𝟬+ 𝘆𝗲𝗮𝗿𝘀 𝗶𝗻 𝗿𝗼𝗯𝗼𝘁 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴, from my PhD at Imperial to Berkeley to building the Dyson Robot Learning Lab, one frustration kept hitting me: 𝗪𝗵𝘆 𝗱𝗼 𝗜 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗿𝗲𝗯𝘂𝗶𝗹𝗱 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗼𝘃𝗲𝗿 𝗮𝗻𝗱

1x's NEO Humanoid has the LOWEST latency VR teleoperation I've ever seen! It matches nearly instantly!

Tired of collecting robot demos? 🚀 Introducing CP-Gen: geometry-aware data generation for robot learning. From a single demo, CP-Gen generates thousands of new demonstrations to train visuomotor policies that transfer zero-shot sim-to-real across novel geometries and poses.

Penetration-free for free, you need OGC. I released the code of our SIGGRAPH 2025 paper: Offset Geometric Contact, where we made real-time, penetration free simulation possible, with Jerry Hsu @zihengliu Miles Macklin Yin Yang Cem Yuksel Page: ankachan.github.io/Projects/OGC/i…

GEM❤️Tinker GEM, an environment suite with a unified interface, works perfectly with Tinker, the API by Thinking Machines that handles the heavy lifting of distributed training. In our latest release of GEM, we 1. supported Tinker and 5 more RL training frameworks 2. reproduced

After carefully watching the video I guess they are transferred mainly through wrist camera (all robots usetvery similar wrist camera setup) along with relative actions (aligned with wrist cam obs)?

🙌

VLAs have become the fastest-growing subfield in robot learning. So where are we now? After reviewing ICLR 2026 submissions and conversations at CoRL, I wrote an overview of the current state of VLA research with some personal takes: is.gd/1pqw9w

With Gemini CLI's new pseudo-terminal (PTY) support, you can run complex, interactive commands like vim, top, or git rebase -i directly within the CLI without having to exit, keeping everything in context.

there’s only one right answer here, the Yann LeCun definition, and everyone should be able to recite it word for word

Introducing RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning. lei-kun.github.io/RL-100/ 7 real robot tasks, 900/900 successes. Up to 250 consecutive trials in one task, running 2 hours nonstop without failure. High success rate against physical

Interesting...

The Andrej Karpathy interview 0:00:00 – AGI is still a decade away 0:30:33 – LLM cognitive deficits 0:40:53 – RL is terrible 0:50:26 – How do humans learn? 1:07:13 – AGI will blend into 2% GDP growth 1:18:24 – ASI 1:33:38 – Evolution of intelligence & culture 1:43:43 - Why self

Very cool work!

I recently realized that the morphology of whole-body robots (beyond table top) shapes how we design teleoperation interfaces — and those interfaces dictate how efficiently we can collect data. So maybe the question isn’t just about better teleop… It’s whether we should design

Mask Robot

Also notice that Tesla has such data on a few types of vehicles, and robot embodiment diverges.