Event

HarnessPAI proposes an evolvable execution framework for physical AI

1 reports · 1 sources

The researchers propose the HarnessPAI framework, which uses code as an executable, evolvable interface that organizes underlying action primitives, and is independent of both model and embodiment. On a desktop robotic arm, a home robot, a robot vacuum, and a bipedal walking agent, the framework achieves consistent improvements over baselines. The framework distinguishes two time scales: within a single rollout, it executes open-loop with a fixed program and checks it; across rollouts, it revises the program in a closed loop based on execution feedback and distills failures into reusable skills. It can also generate training data for tasks in which it is difficult to obtain successful trajectories on its own, for use in fine-tuning VLA and WAM models.