Announcement
DreamCan proposes sparse visual planning for long-horizon robot tasks
Its researchers say it learns from action-free robot videos to select reachable, useful visual subgoals rather than predict every frame.
TLDR
The paper’s authors describe DreamCan as a visual planner paired with a controller that turns selected subgoals into robot actions. They report 91.2% success on LIBERO-Long using a 0.1B-parameter controller, compared with 84.0% for dense-video planning in their tests. They also describe real-world two-arm experiments in box folding and packing.
Combined views
13.7K
6 Sources, first seen ago
