Carnegie Mellon University highlighted Long-Horizon Adaptive Manipulation Planning, or LAMP, on September 24 as a system for coordinating teams of robots that move objects through heavily cluttered spaces. The work was accepted for IROS 2026.
Planning object motion and robot access together
LAMP combines a learned generative model for short-horizon manipulation with search over longer task sequences. Candidate object movements are accepted only after the system checks whether robots can reach the required contacts, navigate without collisions and execute the manipulation. LAMP-A* performs eager search over verified transitions; LAMP-Lazy defers checks, penalizes infeasible edges and replans incrementally.
Promising simulation results, no field validation yet
The researchers evaluated the methods across 100 simulated scenes on four cluttered maps. Their project page reports a median planning time of roughly two to three seconds per segment for LAMP-Lazy, while eager variants often took longer or reached a 500-second did-not-finish threshold. The paper also reports higher success on difficult long-horizon tasks than the tested baselines.
Those results are author-reported and have not been independently replicated for this article. The evaluation is simulation-based, and it does not establish performance with physical robots, imperfect sensing, variable payloads or warehouse safety constraints.