LIMA SPLITS ROBOT "THINKING" FROM "REACTING" INTO TWO ASYNC SYSTEMS, CUTS LATENCY 45.8%

Summary from the TLDR Hardware newsletter

Dexterous manipulation needs both long-term foresight and fast reactive control, but existing models tend to be good at only one, so LiMA decouples the two into an asynchronous dual-system hierarchy: a slow system handles sparse long-horizon intent, a fast system handles dense high-frequency motion, bridged by a Latent Schrödinger Bridge Coupling mechanism that treats aligning the two as a probabilistic transport process rather than a direct mapping. This cuts inference latency by 45.8% versus Cosmos-Policy, and across six bimanual manipulation tasks, LiMA hit a 70.8% overall success rate while holding up on scenarios it wasn't trained on.