SenseNova and its Joint Research Team report that GPT-6-Astra performs well on robotic tasks, but say it has not yet reached practical embodied intelligence. The paper, published on October 15 and submitted to Hugging Face on October 19, evaluates the system across 12 benchmarks covering tabletop manipulation, humanoid robots, mobile platforms, and dexterous hands. Under a shared evaluation protocol, GPT-6-Astra learns from training-set demonstrations provided in context and receives substantially more execution steps. The authors report unprecedented performance under those conditions.

The researchers identify five areas for further work: making in-context learning more general, reducing demonstration costs and execution time, verifying physical states, recovering from irreversible failures, and preserving progress over long tasks. They also find that close-loop coordination between GPT-6-Astra as a high-level reasoner and a low-level controller such as π0.5 is costly but effective.

Real-world experiments using 8 robot configurations expose practical limits: collecting robot demonstrations for in-context learning is expensive, and physical constraints make it implausible to extend execution time simply to raise success rates. The team proposes a prototype for future embodied benchmarks that derives tasks from real-world needs, supplies no in-domain demonstrations, and evaluates agents against a world clock that continues running while they deliberate. The authors conclude that GPT-6-Astra has largely reached embodied competence, but not practical embodied intelligence.

SenseNova and its Joint Research Team report that GPT-6-Astra performs well on robotic tasks, but say it has not yet reached practical embodied intelligence. The paper, published on October 15 and submitted to Hugging Face on October 19, evaluates the system across 12 benchmarks covering tabletop manipulation, humanoid robots, mobile platforms, and dexterous hands. Under a shared evaluation protocol, GPT-6-Astra learns from training-set demonstrations provided in context and receives substantially more execution steps. The authors report unprecedented performance under those conditions.

The researchers identify five areas for further work: making in-context learning more general, reducing demonstration costs and execution time, verifying physical states, recovering from irreversible failures, and preserving progress over long tasks. They also find that close-loop coordination between GPT-6-Astra as a high-level reasoner and a low-level controller such as π0.5 is costly but effective.

Real-world experiments using 8 robot configurations expose practical limits: collecting robot demonstrations for in-context learning is expensive, and physical constraints make it implausible to extend execution time simply to raise success rates. The team proposes a prototype for future embodied benchmarks that derives tasks from real-world needs, supplies no in-domain demonstrations, and evaluates agents against a world clock that continues running while they deliberate. The authors conclude that GPT-6-Astra has largely reached embodied competence, but not practical embodied intelligence.