Sim training in 4.2 GPU-hours teaches robotic hand pen-spinning and cube solving

FinanceYF5 · x · 2026-09-21

FinanceYF5 shares progress on dexterous robotic manipulation: a system referred to as GPT-6 Astra is being trained via large-scale parallel simulation, with zero-shot fine manipulation already impressive.

After training, the robotic hand learns pen spinning, Rubik's cube handling, and hammer use. The cube task used 4,096 parallel simulation environments and roughly 4.2 GPU-hours of training. The stated end goal is sim2real — transferring these skills from simulation into the physical world.

Original post →

More from Embodied

Embodied channel →