Meta's ProWAM hits 70% zero-shot real-world robot success with sparse visual sub-goals
meta · hf · 2026-10-05
Meta introduces ProWAM, a progressive world action model that replaces dense video rollouts with an ordered sequence of sparse visual sub-goals for long-horizon robot control. Sub-goal prediction is learned from action-free videos; a single video-backbone pass caches sub-goal features, leaving only lightweight action denoising at replanning. Results: 85.8% on LIBERO-Plus and 75.7% on randomized RoboTwin (up to +35.9% relative over the strongest baseline), 48.1% on RoboCasa365, and 70.0% zero-shot real-world success in novel scenes vs 55.0% for the best baseline.
More from Embodied
- Galbot's wheeled humanoid G1 now works night shifts at a Shanghai pharmacy — CyberRobooo · 2026-10-05
- Dream4ACT Unifies Video-Action Modeling Across Robot Embodiments, Hitting 89% on RoboTwin 2.0 — Xiangyu Zhu · 2026-10-05
- Munich Robotics Startup RobCo Hits $1B Valuation in Nine Months — lukas_m_ziegler · 2026-10-05
- Physical AI field session in Bengaluru to tackle post-deployment evals and retraining loops — carrycooldude · 2026-10-05
- Meta's NAVA-WAM pretrains robot action policies directly from action-free videos — meta · 2026-10-05
- Robotics' most-hyped model completed just 7 of 100 real manipulation tasks, demo reel hides the rest — carrycooldude · 2026-10-05