Robot whole-body controller training drops from 128 GPUs × 1 week to 1 GPU × 3 minutes
StewartalsopIII · x · 2026-11-15
Robotics developer @makerobotgo reports that before October 20, training a whole-body controller (a robot walking model) took a week on 128 GPUs; now it takes 3 minutes on a single GPU. If accurate, that's orders-of-magnitude gains in embodied AI training efficiency, drastically cutting iteration costs.
More from Embodied
- China unveils autonomous electric passenger jet: 19 seats, 300 km/h, 400 km range — Scobleizer · 2026-11-15
- CoRL 2026 wraps up as General Chair signs off; next stop Paris — yukez · 2026-11-15
- D.C. Council kills robotaxi bill 3-3, labor concerns trump Waymo safety data — QuintinPope5 · 2026-11-15
- Scoop: OpenAI is building factory test infrastructure for robot circuit boards — ryanmerket · 2026-11-15
- SpaceX Appears to Catch an Orbital Upper Stage Mid-Air — VraserX · 2026-11-15
- RAI Institute's RL bicycle robot nails table jumps, flips and wheelies — KyleMorgenstein · 2026-11-15