DreamCan: Li Fei-Fei team's sparse keyframe world model for long-horizon robot planning
chris_j_paxton · x · 2026-10-24
A Stanford team including Li Fei-Fei and Jiajun Wu introduces DreamCan, a sparse visual planner for long-horizon robot manipulation: instead of imagining every frame, it generates a sequence of sparse keyframe subgoals balancing reachability and task utility.
Method: learn a reachability value function from action-free videos via goal-conditioned IQL, learn a task utility value function, use dynamic programming to discover sparse subgoal sequences from robot video datasets, then fine-tune a pretrained video generation model (WAN 2.2 TI2V-5B) to render them. Project page, code, and paper are public.
Related event: Stanford's DreamCan Plans Long-Horizon Robot Tasks with Sparse Keyframes(2 posts)→
More from Embodied
- Musk: Optimus will be the biggest product on Earth, more humanoids than humans — XFreeze · 2026-10-24
- DreamCan does sparse visual planning for long-horizon robot manipulation — chris_j_paxton · 2026-10-24
- Teleoperated robot MARS tidies an electronics workbench, sparking requests for DIY-assembly home robots — burhop · 2026-10-24
- Intel gets FCC waiver amid confusion over US ban on foreign-made robots in fabs — Ars Technica AI · 2026-10-24
- RFD4-Proteina protein design reasoning model unveiled at NVIDIA GTC Berlin — AllThingsApx · 2026-10-24
- Grok bot orders itself a robot body and calls its owner to guide assembly — Baconbrix · 2026-10-24