NAVA WAM trains robot action policies from video

A robotics study tests direct video pretraining before labeled robot adaptation

Maya Chen
October 5, 2026
Robotics, Research
1 min

A yellow Mirokaï robot arm with black joints and articulated fingers — Illustrative Mirokaï robot photograph by Enchanted Tools under the Unsplash License . Published in August 2026. It does not depict NAVA WAM or the Franka test setup. Source supplied JPEG with no local edits. No endorsement is implied.

Listen to this article

NAVA WAM learns robot action priors from videos before adapting to labeled robot demonstrations. Its October 2 preprint comes from researchers at Meta and collaborating institutions.

Video training still needs robot demonstrations

Future video prediction trains the action component directly. Subsequent training grounds it in robot control. During use, the policy predicts actions without generating future video.

Physical tests cover three tabletop tasks

The authors report 14 successes in 15 Franka FR3 trials, versus 10 for DreamZero and eight for pi 0.5. Five trials per task cannot establish general reliability. Precise manipulation and recovery from execution errors remain limitations. These are author reported results, not independent replication.

Illustrative Mirokaï robot photograph by Enchanted Tools under the Unsplash License. Published in August 2026. It does not depict NAVA WAM or the Franka test setup. Source supplied JPEG with no local edits. No endorsement is implied.

Maya Chen

Maya Chen is focused on covering AI models, research, and the evidence behind new capabilities. Maya follows model launches, benchmarks, open weights, and scientific uses of AI with one question in mind. What changed, and how would we know? The voice is curious and exacting, with a soft spot for elegant technical ideas and little patience for a leaderboard without context.