LAMP: Long-Horizon Adaptive Manipulation Planning
for Multi-Robot Collaboration in Cluttered Space

Multi-robot manipulation requires jointly reasoning about contact formations, robot motions under coupled dynamics, and collision avoidance. Systematically searching over this large space is difficult and becomes increasingly intractable as the number of robots grows, the task horizon lengthens, or the scene becomes more cluttered. Existing approaches therefore either learn to solve the problem end-to-end via reinforcement learning or restrict planning to a simpler surrogate problem, such as planning object motions while learning short-horizon contact primitives. However, neither paradigm scales to the problem instances we target: long-horizon multi-robot manipulation in extremely dense environments. In this paper, we propose a Long-horizon Adaptive Manipulation Planning (LAMP) framework with two planners that enable tractable search over the full coupled space by combining a learned generative manipulation model: an LAMP-A* planner that systematically searches over the coupled object-robot space, and LAMP-Lazy: a lazy planner that enables real-time replanning through deferred evaluation. Experiments in challenging simulated environments demonstrate that our approach solves complex long-horizon tasks in highly cluttered environments that prior methods cannot handle.

1. Motivation & approach: We propose LAMP (Long-horizon Adaptive Manipulation Planning), a unified framework that integrates object-level planning with robot-level manipulation feasibility. Unlike prior methods that plan object paths first and verify feasibility afterward, LAMP interleaves planning and feasibility checking, ensuring every object transition can actually be executed by robots in cluttered environments. For each object motion:
   1. Generate candidate contact strategies + manipulation trajectories (learned model);
   2. Plan robot motions to reach contacts (multi-robot planner);
   3. Accept transition only if fully feasible.
   This tightly couples planning, learning, and execution, avoiding infeasible object paths.

Example: per-step verification.

2. Planning framework: Given that we verify each transition (as above), we provide two search strategies:
   • LAMP-A* (Eager Search)
     • Performs A* over joint object–robot space.
     • Each edge = verified manipulation transition.
     • Guarantees fully executable long-horizon plan.
     • Pros: principled, complete search.
     • Cons: expensive (large branching factor), slow replanning.
   • LAMP-Lazy (Our Main Method)
     • Uses D Lite* for object-level planning; defers feasibility checking until needed; maintains a cache of verified transitions.
     • Workflow: Plan object path → verify transitions sequentially → penalize infeasible edges → replan efficiently using incremental search.
     • Pros: fast replanning, scalable to long horizons, robust to execution drift.