Alibaba DAMO's RynnVLA-Latent Hits 98.8% on LIBERO via 3D-Grounded Latent Actions
Alibaba-DAMO-Academy · hf · 2026-10-27
Alibaba DAMO Academy presents RynnVLA-Latent, a framework that turns heterogeneous interaction videos into scalable supervision for robot policy pretraining using 3D-grounded latent actions.
Key points:
- RynnLAM learns continuous latent actions from visual transitions without native robot controls, combining residual future-feature prediction, camera-compensated 3D-flow supervision, and camera-motion decoupling.
- Applied to 30.1K hours of human, robot, and simulation videos, producing 745.52M latent-action descriptions.
- The resulting 4B policy achieves 98.8% success on LIBERO, 84.7% on LIBERO-Plus without benchmark-specific adaptation, 51.9% average across five VLABench tracks, and 88.6% average success on RoboTwin 2.0 bimanual benchmarks.
More from Embodied
- Musk proposes 'real benchmarks' for robots: assembling a Swiss watch while catching a 100mph baseball — elonmusk · 2026-10-27
- Humanoid robots caught 'burning out': photos show them flat on the floor mid-shift — alysha_lobo · 2026-10-27
- Tesla Cybercab line: ~25 stations, 80% automation, no paint shop saves hundreds of millions — mitchdeg · 2026-10-27
- Tesla Tops Waymo in Austin With 438K Weekly Driverless Miles vs 406K — AccBalanced · 2026-10-27
- Tesla FSD safety report: one major collision per 6M miles, 9x safer than US average — elonmusk · 2026-10-27
- Dream Labs is hiring robot learning researchers in SF, plus security and ops roles — scott_e_reed · 2026-10-27