PhysEvo: Astra Can Act, Let It

One-line takeaway

PhysEvo performs physical recursive self-improvement around a frozen Astra model: a task agent runs robot tasks while a meta-agent diagnoses failures and revises tools and skills without weight updates, reaching 62% success on 42 RoboDojo tasks vs 47.17% for a one-shot baseline.

PhysEvo 围绕一个冻结的 Astra 模型做“物理递归自我改进”：任务智能体执行机器人任务，元智能体根据轨迹诊断失败、修改工具和技能，不更新模型权重；在 42 个 RoboDojo 任务上成功率 62%，一次性基线为 47.17%。