Yacine renders 8,000 training worlds at 10M frames/sec on a single GPU

yacineMTB · x · 2026-09-28

YacineMTB details his embodied-AI training pipeline: with custom renderer optimizations, a single GPU renders 8,000 simulated worlds at roughly 10 million frames per second, and near-unlimited worlds can be kept ready to swap in per training episode. He calls it "domain randomization on steroids," noting 8k was simply the batch size that worked best in experiments — a striking example of extreme parallel-simulation optimization for robot policy training.

Related event: Ex-X AI dev claims single GPU renders 8,000 worlds at 10M fps(2 posts)→

Original post →

More from Embodied

Embodied channel →