Mobile Manipulation Planning for Tabletop Rearrangement

TL;DR

This work tackles mobile tabletop rearrangement under object dependencies and limited buffers by extending A*-based planning with multiple relocation actions, region-based manipulation regions, and state re-exploration. It combines explicit/implicit object reasoning, region reduction, and goal-attempting to deliver high-quality plans faster than prior planners such as STRAP, TRLB, ORLA*, and MCTS, while remaining an anytime method. Key contributions include a shortening/rewriting mechanism to merge action sequences, a region-aware successor generation, and a greedy goal-attempting loop that yields feasible plans early and improves over time. The proposed approach demonstrates substantial improvements in solution quality and planning time, highlighting its practical impact for scalable mobile-robot tabletop rearrangement, though optimality cannot be guaranteed.

Abstract

Efficient tabletop rearrangement planning seeks to find high-quality solutions while minimizing total cost. However, the task is challenging due to object dependencies and limited buffer space for temporary placements. The complexity increases for mobile robots, which must navigate around the table with restricted access. A*-based methods yield high-quality solutions, but struggle to scale as the number of objects increases. Monte Carlo Tree Search (MCTS) has been introduced as an anytime algorithm, but its convergence speed to high-quality solutions remains slow. Previous work~\cite{strap2024} accelerated convergence but required the robot to move to the closest position to the object for each pick and place operation, leading to inefficiencies. To address these limitations, we extend the planner by introducing a more efficient strategy for mobile robots. Instead of selecting the nearest available location for each action, our approach allows multiple operations (e.g., pick-and-place) from a single standing position, reducing unnecessary movement. Additionally, we incorporate state re-exploration to further improve plan quality. Experimental results show that our planner outperforms existing planners both in terms of solution quality and planning time.

Figures (12)

• Figure 1: A mobile robot performs tabletop rearrangement, needing to move its base to access the entire table due to limited reach. Blocks of various colors are scattered and must be relocated to matching color positions.
• Figure 2: (a) The current state's arrangement is represented by solid circles and goal arrangement indicated by a dashed circle. (b) Picking regions. For each object, the highlighted table side in the same color indicates where the robot can stand to pick it up. When regions overlap, the colors are evenly divided, with each occupying half of the space. (c) Placing regions. For each object, the region with the same color on the table side indicates the region where the robot can stand to place.
• Figure 3: The process of generating manipulation regions, along with their corresponding object-operation dictionary defining the feasible operations on objects within each region. Left: Based on placing and picking regions, find initial manipulation regions with their object-operation dictionary on one table side, represented as a set of indexed green boxes. Right: The filtering process removes non-manipulable objects from the object-operation dictionary and eliminates regions if they become empty. The merging process then combines the filtered regions.
• Figure 4: The left-bottom box is the robot. (1) In a given state, there are two manipulation regions, R1 and R2. For R1, points $a$ and $c$ are efficient standing locations. For R2, only point $c$ is efficient because it lies on the shortest path from the robot to point $b$. (2) When the object is in hand and the placing region is on the right side, only point e is the feasible placing point. (3) When the object is in hand and the placing region is on both the top and right side, both f and g are feasible placing points.
• Figure 5: Given a state where the robot only moves on one table side, there are 4 successor states, shown in the bottom, lead by 4 feasible operation sequences from the current state.