Test‑time Compute Boosts Robot Policy General
Today (June 3), I'll be speaking at CVPR at the Test-Time Scaling for Computer Vision WS (1:30 pm PT) about how we can use test-time compute to boost generalization of robot policies, room 506. Also speaking *right now* (in 5 min) in the Deployment of Foundation Models WS in Ballroom 2!
Rethinking What Makes a Good World Model
Tomorrow in the Workshop on World Models at ICLR in Rio (10:30 am) I’ll talk about a… different take on what might make for a good world model. Come find out, 10:30 in Room 202A at ICLR https://t.co/TRzkSqeeKN
RL Boosts Robot Foundation Models; Generative AI Enables Self‑Improvement
I'll give a talk about lifelong learning tmrw (Sun), 9:30 am (Brazil time) in the lifelong agents workshop, about how we can get robot foundation models to improve with RL: https://t.co/nKci27JeaH Then at 11:30 am, I'll talk about how generative models...
π0.7 Demonstrates Emergent Compositional Generalization via Instructions
We finished evaluating π0.7, our new model at Physical Intelligence. What I'm most excited about with π0.7 is that it's starting to show some surprising emergent compositional generalization, being able to both perform complex tasks and learn new tasks just...
Pi Models Power Unexpected Flying Gripper Drones
Didn't think that "drones with grippers" was in the cards for a likely embodiment for pi models, but there we have it. It's literally a flying gripper. But believe it or not, pi models have been used on even stranger...
Structured Models Generate Higher‑Reward Material Designs
A while ago we figured out that structure enables data-driven design: if we have data of designs + rewards, we can find a design with *higher* reward if we learn a structured function: https://t.co/HlevRSTMXV In our latest work, @kuba_AI developed a...
Fast Online RL Upgrades Π-06 with 15‑minute Data
Back in Nov we developed Recap and trained π*-06 with RL. Now, we developed a fast *online* RL method that improves π-06 with as little as 15 min of robot data for precise tasks, using "RL tokens" exposed by our...
Multi-Scale Embodied
We made a memory system for our models at PI. We call it Multi-Scale Embodied Memory (MEM). It provides both short-term and long-term memory to enable very long tasks. We tested it on cleaning a kitchen (and yes, washing dishes),...
Run VLAs in Real Time Using Fast Edge Adapter
Check out Noriaki's thread about a way to get VLAs to run in real time with a fast "edge adapter"!
VLA Reasoning Lets Vehicles Safely Navigate Complex Edge Cases
VLAs can enable vehicles to better handle complex edge cases: a VLM can "think through" a complex interaction, deduce a common sense behavior, and then a VLA can carry that out to maintain safe(r) behavior even in unusual situations.