GPT-6-Astra nears saturation on 12 robot benchmarks but burns 9M tokens per task
rohanpaul_ai · x · 2026-10-22
A joint team from SenseNova, Ropedia, NTU Singapore and the University of Tokyo evaluated GPT-6-Astra across 12 robot benchmarks spanning tabletop manipulation, humanoids, mobile platforms and dexterous hands.
Key findings
- With in-context learning from training demos and up to 10x the usual action steps, GPT-6-Astra neared saturation on the LIBERO suite and hit 81.9% on RoboDojo and 87.3% on DexJoCo.
- The cost is prohibitive: a single LIBERO-Ultra-Long episode used 9M tokens and nearly 2 hours; on real robots it could not trade more steps for higher success, verify physical outcomes, or recover from irreversible mistakes.
- The team also found closed-loop coordination between a high-level "System 2" reasoner and a low-level "System 1" controller is costly but effective, and prototyped next-gen embodied benchmarks with real-world tasks, no in-domain demos, and a continuously running world clock.
Conclusion: GPT-6-Astra has largely crossed the threshold of embodied competence but not yet practical embodied intelligence — general models work as robot controllers for research, while deployment remains blocked by demo cost, compute, and physical limits.
More from Embodied
- Tuesday Labs shows off Voxel, a robot butler starting out by cleaning crumbs — CurieuxExplorer · 2026-10-22
- Musk: Tesla will carry the majority of passenger miles on Earth wherever allowed — alexmacgregor__ · 2026-10-22
- Flying a drone with Codex: 12–20s per decision is the bottleneck — kfountou · 2026-10-22
- Robotics fashion is now an industry, observer says — micoolcho · 2026-10-22
- Close-up photos from a Chinese expo reveal a humanoid robot's joint and torso design — micoolcho · 2026-10-22
- Eric Jang: robots need open SDKs so the community can secure them at scale — ericjang11 · 2026-10-22