SeeQ ablations: TD learning, VLM pretraining and language subtask decoding all matter
aviral_kumar2 · x · 2026-09-29
Aviral Kumar shares SeeQ ablation results highlighting three key ingredients: TD learning done the best-of-N way; both pretrained VLMs and pretraining on robot data; and actually decoding the active subtask in language. Further diagnostics of the learned value function are in the paper.
Related event: CMU's SeeQ: Generalist Robot Value Functions Nearly Double Success Rates(3 posts)→
More from Embodied
- Hopping Drone at IROS 2026 Is Fun to Watch, Use Case Unclear — jonstephens85 · 2026-09-29
- Meta's Muse Charm faces iKairos, whose LingOS sees and hears users for deeper context — eyishazyer · 2026-09-29
- Chestnut unveils 18-DOF Aero Hand plus matching exoskeleton for zero-gap humanoid data capture — chris_j_paxton · 2026-09-29
- Flourish's $3,555 wheeled home robot FLOURISH 1 learns tasks from a single demo — CyberRobooo · 2026-09-29
- AMD ships Ryzen AI Max+ PRO with 192GB unified memory, 50% more than NVIDIA's upcoming Spark — ryanshrout · 2026-09-29
- Real-world test: driving on grass takes roughly 3x the torque of concrete — _Stocko_ · 2026-09-29