Generative Trajectory Planning: robots learn to plan with world models instead of RL
du_yilun · x · 2026-10-24
Google researcher Yilun Du introduces Generative Trajectory Planning (GTP), an approach to effective online learning for embodied agents. Instead of learning policies via RL, GTP learns to propose promising action sequences and refines them with a learned world model, improving its planning through experience.
The method shows strong online decision-making on manipulation tasks where RL has struggled. Work led by Haldun Balim with Lina Na Li.
More from Embodied
- Robotics model leaderboards are premature, argues practitioner: task-specific evals first — paigeinsf · 2026-10-24
- Claude designs a 200 lb-payload hexapod robot in one hour, with parts list — _Stocko_ · 2026-10-24
- iPhone Duo hands-on: clunky as a phone, but unfolds into a perfect iPad Mini — firstadopter · 2026-10-24
- Rice & NVIDIA's Sensorimotor Alignment predicts robot policy success without training (r=0.717) — CSProfKGD · 2026-10-24
- Hugging Face adds OpenUSD support as NVIDIA drops 2,500+ free SimReady 3D assets — Promptmethus · 2026-10-24
- NVIDIA's Robot Ducks Steal the Show in Charming Demo — johnkwaters · 2026-10-24