Report: Astra 6.2 Leads Robotics Capability but Least Likely to Aid Autonomous Weapons
RishiBommasani · x · 2026-11-20
Rayan Krishnan shares his team's new report finding that Astra 6.2 is the most capable model for robotics tasks while also scoring lowest on likelihood of aiding autonomous weapon operation. The team disclosed findings to labs two weeks before public release, giving them lead time to curtail this new category of misuse.
The core argument: the most effective approach is for safety testing to keep pace with the frontier rather than inhibit progress; capability measurements can predict prospective safety risks and enable mitigation as fast as models improve. The author notes going live on the chosen timeline was not an easy decision, and thanks Rishi Bommasani, ddkang and Jacob Steinhardt for helping reason through the course of action.
More from Embodied
- Robotics startup Hummingbird hosts first Istanbul event, hiring engineers — animesh_garg · 2026-11-21
- OpenAI NYC hardware hackathon: Corner Walk turns your GPS-tracked stroll into a retro game — OpenAIDevs · 2026-11-21
- OpenAI hardware hackathon: attendees build a ChatGPT-powered Tamagotchi, voice input falls short — OpenAIDevs · 2026-11-21
- Robotics researcher: extending supermimic to manipulation is hard without low-dim collision geometry — KyleMorgenstein · 2026-11-21
- 3DGS vs Point Cloud comparison looks great; surface reconstruction for sim collisions next — jonstephens85 · 2026-11-21
- Haozhi Qi lectures on building dexterity with simulation, human data and tactile sensing — YuXiang_IRVL · 2026-11-21