RoboTok: Retrieving Internet Videos for Robot Learning, +42.2pp Task Success
ceciletamura · x · 2026-10-15
Rice University researchers introduce RoboTok, a scalable data engine that retrieves manipulation-relevant demonstrations from internet videos to train dexterous robot policies. It learns a latent motion space from 3D hand trajectories in actor-centered frames, making retrieval robust to viewpoint, scene, and occlusion changes. In real-world robot experiments, RoboTok-guided policies improve mean task success by 42.2 percentage points over the strongest per-task baseline. (arXiv:2609.03199)
Related event: Rice Team Releases RoboTok, Mining Internet Videos for Robot Demonstrations(2 posts)→
More from Embodied
- Tesla adds 30 more Cybercabs in Texas, purpose-built fleet hits 349 with 300+ serving Austin — XFreeze · 2026-10-15
- Physical AI webinar Oct 22: four robotics founders on evals, retraining, and field failures — carrycooldude · 2026-10-15
- Vela: spline-based continuous trajectories scale vision-language-action models — Yifan Li · 2026-10-15
- Scoble: healthcare robots arrive next year, true humanoids likely 2-7 years away — realmeetjames · 2026-10-15
- KAIST's BeetleBot: finger-controlled soft robot walks, grips and streams sensory feedback — CurieuxExplorer · 2026-10-15
- 14-year-old builds handheld AI pesticide scanner PestiSCAND with 85% accuracy — nordicinst · 2026-10-15