Light-O1 puts 79.3% on simulated robot tasks

Light Origins released Light-O1-Preview, a public text-to-action model that reports 79.3% macro success across 24 simulated RoboCasa humanoid tasks. Its release write-up says evaluation used 50 episodes per task, while the Apache-2.0 model card outputs 138 action values per frame at 20 FPS; the benchmark is simulated and the preview still needs a separate low-level controller. Robotics teams should treat it as an inspectable pretraining direction, not a hardware-ready switch, alongside the archive’s long-horizon agent reliability benchmark.