PhysEvo turns frozen Astra into robot learner.

PhysEvo’s strongest idea is that embodied self-improvement can happen in the harness rather than inside the model. That makes improvements inspectable and reusable, but puts the burden on reliable failure diagnosis and regression testing.

• –RoboDojo success reaches 62%, up from 47.17% for the one-shot Astra baseline.
• –On eight difficult manipulation tasks, PhysEvo improves success from 1.25% to 55%.
• –Simulation-evolved skills transfer to an AgileX PiPER robot, achieving 84% success across five physical tasks.
• –The recursive meta-agent can improve its own diagnostic tools, creating a feedback loop beyond ordinary prompt or skill refinement.
• –The approach is promising for developers who need adaptable robotics systems without expensive policy retraining, though results remain task- and harness-dependent.