Topic

Simulation for robot learning

Physics-based simulators and synthetic data pipelines used to train and validate robot policies before they run on hardware, and the sim-to-real gap that limits them.

Current clusters

build2 publishers

A per-step 66 percent has to compound across the dozens of steps in Skild's S1 tasks

Skild says a single video demonstration teaches its S1 model an unseen ten-minute task with no weight updates. The accuracy figure published alongside that work is 66 percent per manipulation step.

Publishers:nvidianews.nvidia.comruntimewire.com

Reality

Evidence34
Adoption55
Hype gap+32
Incentives82
Confidence60