01/04
Research · Embodied LLM agents
Auto Adapter.
Physics-in-the-loop robot driver synthesis from MJCF/URDF for tool-using agents.
Problem → Build
Tool-using LLM agents usually assume the robot already exposes a Python driver: move_cartesian, get_ee_pose, gripper_close. Auto-Adapter gives the agent that missing layer too: read MJCF/URDF, synthesize a reusable driver, and revise it against MuJoCo physics until the harness passes.
Architecture
FIG. 01a
Five-phase driver synthesis · physics-in-the-loop
Auto-Adapter is a five-phase LLM-agent pipeline: Study reads the robot spec, Generate writes the driver, Validate re-checks it in a fresh MuJoCo simulator, Export bundles a reusable API, and Demo lets a downstream agent call it. The key verifier is not text output but post-step physics state; across eight robots, onboarding averaged $2.98 and 7.8 minutes.
