RoboArchitect automatically redesigns the control code around a frozen robot policy, and it does this without updating any model weights.
How it works: A coding model watches the robot's rollouts, diagnoses failures, and revises the harness, the control code that decides when to query the policy, consult vision tools, or command the arm directly.
Results: Average success on unseen scenes rises from 9.7% to 43.3% (4.5x), and on unseen tasks from 12.7% to 40.2% (3.2x), across multiple benchmarks and policies.
Memory: The harness also evolves a persistent memory, which the paper says can carry learned details such as effective grasp depths into later decisions.
Availability: Code and demos are at roboarchitect-ai.github.io.
Sign in to comment.
TL;DR of the paper, generated automatically.
RoboArchitect automatically redesigns the control code around a frozen robot policy, and it does this without updating any model weights.