Coding Agents for Generalized Task and Motion Planning

Meta description: Astra robot policy evaluations grow from one-off demos to 98,000 tests across 28 simulated environments, with access rules and physical reasoning.

Tags: Astra robot policy, simulated environments, coding agent, physical reasoning, 98,000 evaluations

Featured image: Abstract technology style, no real people, no company logos, no text: glowing nodes connected to a robot arm silhouette in a simulated environment grid.

Astra robot policy evaluations moved from isolated demos to a systematic test program: 98,000 evaluations across 28 simulated environments, with clear access rules and physical reasoning checks.

Early demos

Astra appeared in robotics demos centered on asking a coding agent to write a robot policy. One demonstration cannot show reliability, so the next step was a repeatable test design.

Months of setup

Over months, the effort became a test program of 28 simulated environments. Each task uses access rules, letting evaluators see whether generated policies stay within allowed actions.

Latest results

The latest stage includes 98,000 evaluations. That scale moves beyond anecdotes and exposes both failures and surprisingly clever physical reasoning.

Current stage

This is an evaluation milestone, not a final robotics product. The next signal is whether these simulated results generalize to harder tasks and wider access rules.