TL;DR of the paper, generated automatically.

A new benchmark tests whether coding agents can control robots.

• Control: agents pick direct control, write code, or train a policy.
• Scale: 1,156 tasks, 35 suites, 47 robots, 9 suites are new.
• Result: best of eight agents, GPT-6 Astra in Codex, hits 34% in Standard Setting.