RLinf + SGLang runs Cosmos3 at 3.33x end-to-end eval throughput for robots

ying11231 · x · 2026-09-21

LMSYS's new blog details RLinf, an open-source embodied AI framework, now supporting Cosmos3 from fine-tuning to robot evaluation with 3.33x end-to-end evaluation throughput via SGLang. SGLang batches inference for 128 parallel environments on 8 GPUs covering 500 LIBERO-10 episodes, while RLinf overlaps CPU simulation with GPU inference to cut stage waiting.

Original post →

More from Embodied

Embodied channel →