GPT-6 Astra driving a walking robot: 30s of walking takes 250 calls and 2.8 hours
BootNew7978 · reddit · 2026-11-13
A Reddit post cites the Galbot team's paper "Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies" (arXiv 2609.38537): driving a walking robot directly with GPT-6 Astra, one 30-second simulated locomotion run took 250 model calls at 39.86s each — about 2.8 hours total, with physics paused while the model thinks.
Takeaway: multimodal LLMs reason well, but as an embodied control layer the latency and call overhead remain prohibitive — direct robot control still calls for dedicated robot models like LingBot-Touch. On the same time scale, the 30-second walking bar is a thin line at the left edge.
Related event: Galbot Tests GPT-6 Astra Robot Control: 30-Second Walk Takes 2.8 Hours(2 posts)→
More from Embodied
- Open robot learning book adds planning chapter covering CEM and loop control — GlenBerseth · 2026-11-13
- Gazebo community meeting: how OTTO by Rockwell validates industrial AMRs — ___Mufasaa · 2026-11-13
- 2.8 hours of inference for 30 seconds of walking: GPT-6 Astra as a robot policy tested — Willing_Amphibian825 · 2026-11-13
- Origami Robotics debuts Gen2 full direct-drive robot hand at CoRL dexterity night — viddivj · 2026-11-13
- Fable 5.2 hits 64.9% on RoboDojo; can frontier models replace dedicated robot policies? — Buffalo4516 · 2026-11-13
- Nuro Driver demos fully autonomous rides in SF on Lucid vehicles via Uber — R_D · 2026-11-13