DreamCan does sparse visual planning for long-horizon robot manipulation
chris_j_paxton · x · 2026-10-24
DreamCan (led by Chongkai Gan) tackles long-horizon robot manipulation with sparse visual planning: instead of imagining every frame, the robot reasons over a few sparse visual frames that are both reachable by the robot and make progress toward the task goal.
Key points:
- The capability is learned from action-free, suboptimal videos — no action labels needed
- Framed as an answer to "AlphaGo and coding agents need reasoning; robotics is orders of magnitude harder — what's the right representation?"
- Project page, code, and paper are all publicly released
Related event: Stanford's DreamCan Plans Long-Horizon Robot Tasks with Sparse Keyframes(2 posts)→
More from Embodied
- Princeton's AIM cuts world model training data needs by up to 85x via action influence — shahdhruv_ · 2026-10-24
- NVIDIA expands Cosmos with new tools as Porsche, Mercedes, Siemens adopt physical AI — NVIDIAAI · 2026-10-24
- Enact exits stealth with $30M pre-seed, Switzerland's largest, sensing touch via robot motors — lukas_m_ziegler · 2026-10-24
- Robotics APIs Emerge: Generalist and Physical Intelligence Open Their Best Models to Startups — binarybits · 2026-10-24
- Musk: Optimus will be the biggest product on Earth, more humanoids than humans — XFreeze · 2026-10-24
- Teleoperated robot MARS tidies an electronics workbench, sparking requests for DIY-assembly home robots — burhop · 2026-10-24