Peking University's LiMA uses async diffusion world model to cut robot latency

2C_ornot2C · x · 2026-11-10

Peking University and BAAI introduce LiMA, a bimanual dexterous manipulation framework that reduces real-time latency via asynchronous computation:

Future Dreamer: a 2B-parameter diffusion transformer initialized from NVIDIA Cosmos-Predict2-2B, fine-tuned per task, imagining future multi-view visual latents plus a coarse action prior — runs only once every four cycles.
Motion Refiner: a 1B-parameter DiT trained from scratch, running every step, conditioning on the imagined intent, live multi-view images and proprioception to emit 32-step action chunks.
An Image-to-Image Schrödinger Bridge connects the two, transporting the Dreamer's coarse intent to grounded actions.

The slow-imagine / fast-refine split is a notable new approach to cutting robot inference latency.

More from Embodied