NAVER's WATCH: Cross-Video Human-Video Pretraining Lifts Robot Success Rate to 77.7%
naver-ai · hf · 2026-10-14
NAVER AI Lab introduces WATCH, a cross-video pretraining method for world-action models. Instead of within-video next-frame prediction — which lets models lean on shared appearance — WATCH trains on task-paired human demonstrations across diverse environments: given one demonstration's video and actions, the model jointly predicts another demonstration's actions and future frames from its current observation. This forces learning of task-relevant interactions that transfer across scenes and demonstrators, scalable with human-human pairs alone, no matched human-robot data needed.
After robot post-training, MimicDroid in-context success jumps from 40.9% to 68.3% with Cosmos Policy and from 59.8% to 77.7% with Cosmos3, with the largest gains on unseen levels; gains persist across embodiments and beat within-video and co-training baselines.
More from Embodied
- SlashVibe, a physical controller for AI agents, hits Kickstarter October 27 — sujingshen · 2026-10-14
- Fchollet: Autoregressive Sampling Alone Can't Solve Real-World Robotics — fchollet · 2026-10-14
- CUHK's Orchestra robot framework completes 7.35-step subtask sequences — CUOHK · 2026-10-14
- Robot Learns Hand Skills From 1M Hours of First-Person Video, Generalizes to Unseen Tasks — Scobleizer · 2026-10-14
- Nvidia's Octi shifts dexterous manipulation to sim-based RL, academia rejoices — KyleMorgenstein · 2026-10-14
- Natura-1 robot demos six real-world tasks with sustained autonomous execution — stingning · 2026-10-14