A robotics planning system called MOSAIC reframes long-horizon manipulation by asking which skills will work where, not just which skills exist.
Developed by researchers and detailed in a new paper, MOSAIC uses physics simulation to identify what the authors call "islands of competence" — regions of a task where a given skill reliably executes. Two complementary components handle the work: Generators map those high-confidence zones, while Connectors stitch the resulting trajectories together by solving boundary value problems between them. The system works with a heterogeneous skill set that includes diffusion models, motion planning algorithms, and manipulation-specific models — meaning it is not locked to any single control paradigm. Tests ran in both simulation and on physical hardware across complex, multi-step tasks.
Most planners either search a combinatorially large space of skill sequences and get lost, or compress the problem into symbolic representations that are faster to search but brittle and expensive to hand-engineer. MOSAIC sidesteps both traps by grounding planning in simulation-estimated outcomes, letting the robot focus effort where success is actually probable. For robotics labs trying to get general-purpose arms to handle novel tasks without per-task programming, that is a meaningful shift in approach.
The catch, as with most sim-to-real robotics work, is whether physics simulation stays accurate enough to trust when skills interact in unexpected ways — a gap the paper demonstrates progress on but does not fully close.