GPT-6 Astra aces spatial reasoning benchmarks, beating MolmoAct2

openai's gpt-6 astra was tested against ai2's molmoact2 in a robotics benchmark called stationerybench, which involves five desk-object tasks. gpt-6 astra controlled a dual-arm yam robot and completed 7 out of 100 tasks, with a median progress score of 46. in contrast, molmoact2 completed zero tasks with a median score of 12. ai researcher yoav artzi calls astra a step change in spatial reasoning, with accuracy close to human level on the remap benchmark. this achievement may be attributed to openai's use of large amounts of 3d data, such as blender scenes, in training gpt-6 astra.