GPT-6 Astra Directly Controls Quadruped Locomotion

After researchers questioned whether LLMs can directly act as quadruped locomotion policies, GPT-6 Astra was tested controlling a simulated quadruped at 50Hz joint targets, successfully walking using about 250 inference calls for 5 seconds.

2026-09-09 ~ 2026-09-09 · 2 related posts

• GPT-6 Astra as a Quadruped Locomotion Policy: 250 Inferences, 5 Seconds of Walking — YuXiang_IRVL · 2026-09-09
• GPT-6 Astra implements a quadruped locomotion policy directly, stretching the 'LLM' label — YuXiang_IRVL · 2026-09-09