GPT-6 Astra Tops Independent Robot Manipulation Eval at 70.5%
Loading video
Loading videoOn 200 episodes drawn from an Allen AI benchmark, GPT-6 Astra reached 70.5% success in independent robot manipulation tests, ahead of Cosmos3 (53.0%), MolmoAct-2 (38.6%) and pi0.5 (23.2%). Single-arm pick-and-place hit 49/50 in simulation, while bimanual packing and clothing folding stayed at 26%.
Category: research
Author: @yokoshaan
Date: 2026-09-21T00:00:00
Duration: 7.2s
