GPT-6 Astra shows early gains in robotic spatial reasoning
OpenAI’s GPT-6 Astra recorded stronger results than Ai2’s MolmoAct2 in StationeryBench, a new robotics benchmark focused on interactions with objects on a desk. The evaluation covered five tasks, including uncapping a marker, emptying paper clips and passing a ruler between two robotic arms. Both systems controlled identical dual-arm YAM robots in 200 trials.
Astra fully completed seven of 100 tasks, while MolmoAct2 completed none. Its median progress score was 46 out of 100, compared with 12 for MolmoAct2. The benchmark’s results, videos and code have been published on GitHub. Cornell and Google DeepMind researcher Yoav Artzi described the performance as a “step change” in spatial reasoning. On the still-unpublished REMAP benchmark, he said GPT-Astra’s accuracy was close to human performance, while remaining below people in other situations. Artzi speculated that extensive training on 3D material, including Blender scenes, could help explain Astra’s improvement on three-dimensional tasks. OpenAI has long-term plans to develop consumer robots.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.