PhysEvo: Astra Can Act, Let It

PhysEvo. Astra Can Act, Let It. By Wenqing Tian and five others, from the University of Chinese Academy of Sciences, the Institute of Automation of the Chinese Academy of Sciences, and two other institutions. Published on the sixth of October.

The title makes the opening claim: the model Astra can act. What holds it back is not the model itself. Reliable manipulation depends on the system through which a model observes and controls the world, the tools and skills wrapped around it. If that surrounding system is weak, a capable model still fails at the task.

PhysEvo is a framework for physical recursive self-improvement, built around a single frozen model. Rather than training the model, it improves the system around it, and it lets that improvement compound. Recursive is the key word. The process that does the improving is itself one of the things that gets improved, so each round can leave the system better at acting and better at working out how to act better.

Two agents divide the work. A task agent executes robot tasks. A meta-agent takes the resulting trajectories, diagnoses failures, revises tools and skills, and tests the corrections. The meta-agent can also improve its own diagnostic tools, so revisions that are kept support later action and later self-improvement alike. Through this process the system develops joint-level control, observation that seeks out evidence, and reusable manipulation skills, with no updates to model weights and no separately trained action policy.

Across forty-two RoboDojo tasks, retained task-specific deployment versions were evaluated on held-out layouts. They score sixty-eight point one four out of a hundred on a five-dimension average, with sixty-two percent success. The comparison is RoboDawn's one-shot Astra agent at forty-seven point one seven percent, the strongest published reference the authors compared against.

On eight manipulation tasks that challenge Astra used directly, PhysEvo reaches fifty-five percent success, against one point two five percent for the direct-Astra reference. The harness evolved in simulation was then deployed on AgileX PiPER, with skill revision continuing there. Across twenty-five trials on five real-world tasks, it scores ninety point six out of a hundred on average, with eighty-four percent success.

The phrase the authors use for the whole loop is that PhysEvo turns the consequences of action into persistent, testable changes to how a frozen model acts and improves. The model never changes. Everything that gets better lives in the harness, as tools, skills and diagnostics that persist from one attempt to the next and can be tested before they are kept.