During long-horizon task planning, a robot must decide what to do, while ensuring each action is geometrically feasible.

To this end, we propose 𝗽𝗹𝗮𝗻𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗶𝗻𝘁𝗲𝗿𝗹𝗲𝗮𝘃𝗲𝗱 𝘃𝗶𝘀𝗶𝗼𝗻-𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝘁𝗵𝗼𝘂𝗴𝗵𝘁𝘀.

🤔 Can VLA models teach themselves the best way to reason?

We introduce R&B-EnCoRe which is a self-improving pretraining cycle for reasoning VLAs to propose and self-discover what embodiment-specific reasoning looks like.
🔗 milanganai.github.io/rnb-encore
📃 arxiv.org/pdf/2602.08167

test-time compute [ttc] in robotics isn't free & isn't always worth it. smart allocation of ttc recovers frontier-level planning at a fraction of the cost!

coauthor @milanganai
w/ Yasmina @ajaysridhar0 Mozghan @katielulula Clark Barrett @jiajunwu_cs @chelseabfinn @drmapavone 🧵

4/4 we are incredibly excited about this line of work!! feel free to reach out to us 😁

📄 arxiv.org/abs/2606.12402
🔗 jadee-dao.github.io/direct/
🤖 jadee-dao.github.io/direct/video-d…