PhysEvo: Astra Can Act, Let It
PhysEvo reports large manipulation gains by letting a frozen Astra model revise its tools and skills through robot-task failures.
The framework uses a task agent to act and a meta-agent to diagnose trajectories, change tools and skills, and test fixes. It keeps those revisions without updating model weights or training a separate action policy. Across 42 RoboDojo tasks, retained deployment versions reached 62.00% success, versus 47.17% for the cited RoboDawn one-shot Astra baseline. On five real-world AgileX PiPER tasks, the simulation-evolved harness plus continued skill revision reached 84.00% success across 25 trials. HF Daily Papers' note
The framework uses a task agent to act and a meta-agent to diagnose trajectories, change tools and skills, and test fixes. It keeps those revisions without updating model weights or training a separate action policy. Across 42 RoboDojo tasks, retained deployment versions reached 62.00% success, versus 47.17% for the cited RoboDawn one-shot Astra baseline. On five real-world AgileX PiPER tasks, the simulation-evolved harness plus continued skill revision reached 84.00% success across 25 trials. HF Daily Papers' note
score 4