Stanford's RoboArchitect lifts frozen robot policies from 9.7% to 43.3% success without touching weights
dair_ai · x · 2026-11-03
Stanford's RoboArchitect paper shows harness optimization as an alternative to finetuning: a coding model watches robot rollouts, diagnoses failures, and iteratively rewrites control code around a frozen policy—deciding when to query the policy, use vision tools, or command the arm directly. Without changing a single weight, success on unseen scenes rises from 9.7% to 43.3% (4.5×) and on unseen tasks from 12.7% to 40.2% (3.2×). Gains are largest far from the policy's training data, where a general-purpose policy inside the evolved harness often beats its finetuned version. The search also discovers intuitive routines like lifting a jammed arm clear, and the harness evolves persistent memory such as effective grasp depths.
Related event: Stanford's RoboArchitect Boosts Robot Success Without Touching Weights(3 posts)→
More from Embodied
- RoboColosseum assembles robot fleet to benchmark AI models in the real world, live in 48 hours — DJiafei · 2026-11-03
- Yacine's drones keep crashing into windows like real birds, and vision-only fix is unsolved — yacineMTB · 2026-11-03
- Scoble sides with call for VR talent to flee to robotics after '10 years getting kicked in the face' — Scobleizer · 2026-11-03
- Claude Opus delivered 50 3D-printed and 210 hardware parts — and it works — IMordatch · 2026-11-03
- Neuralink user navigates hospital to radiology using only thought control — GregCook2011 · 2026-11-03
- Robotics team to unveil system at RoboColosseum in 48 hours, demos era over — DJiafei · 2026-11-03