ReloPush-BOSS: Optimization-guided Nonmonotone Rearrangement Planning for a Car-like Robot Pusher

TL;DR

ReloPush-BOSS tackles nonmonotone multi-object rearrangement in dense clutter using a car-like pusher by embedding prerelocation optimization into a Push-Traversability graph and guiding prerelocations with Dubins path insights. It combines a depth-first sequence planner with an optimization-based prerelocation search and seed-based warm starts to avoid high-cost local minima, achieving scalable planning for up to 13 objects. Empirical evaluation shows higher success rates, shorter pushing paths, and competitive planning times compared to baselines, with successful real-robot demonstrations on a 1/10 scale racecar. This approach advances robust nonprehensile rearrangement planning in constrained, realistic settings and provides a practical pipeline for complex pushing tasks.

Abstract

We focus on multi-object rearrangement planning in densely cluttered environments using a car-like robot pusher. The combination of kinematic, geometric and physics constraints underlying this domain results in challenging nonmonotone problem instances which demand breaking each manipulation action into multiple parts to achieve a desired object rearrangement. Prior work tackles such instances by planning prerelocations, temporary object displacements that enable constraint satisfaction, but deciding where to prerelocate remains difficult due to local minima leading to infeasible or high-cost paths. Our key insight is that these minima can be avoided by steering a prerelocation optimization toward low-cost regions informed by Dubins path classification. These optimized prerelocations are integrated into an object traversability graph that encodes kinematic, geometric, and pushing constraints. Searching this graph in a depth-first fashion results in efficient, feasible rearrangement sequences. Across a series of densely cluttered scenarios with up to 13 objects, our framework, ReloPush-BOSS, exhibits consistently highest success rates and shortest pushing paths compared to state-of-the-art baselines. Hardware experiments on a 1/10 car-like pusher demonstrate the robustness of our approach. Code and footage from our experiments can be found at: https://fluentrobotics.com/relopushboss.

Figures (10)

• Figure 1: We introduce ReloPush-BOSS, an optimization-based planning framework for nonprehensile multiobject rearrangement planning using a car-like robot pusher. By capturing kinematic, geometric, and physics constraints into a unified object traversability graph (\ref{['fig:relopush_before']}), our framework is capable of handling nonmonotone instances in densely cluttered workspaces with up to 13 objects (\ref{['fig:relopush_after']}).
• Figure 2: ReloPush-BOSS embeds a PT-graph in a depth-first high-level planner. Starting from the initial configuration (d0), it builds a PT-graph at each depth, enumerates candidate rearrangements, and orders them by cost in a priority queue (ordered candidate queue). Candidates are tested in order; infeasible ones (e.g., inaccessible objects) are pruned and the planner backtracks until it finds a feasible sequence or exhausts all options.
• Figure 3: Push-Traversability graph (PT-graph). Vertices denote pushing poses for objects/goals; directed edges encode feasible pushes weighted by total pushing length. (a) Pushing poses are defined as contact normals on object's sides. (b) Workspace view showing traversable connectivity between objects/goals. (c) Corresponding graph; for clarity, only one vertex per object is shown.
• Figure 4: Illustration of a prerelocation. (\ref{['fig:prereloa']}) Top: A continuous push-transfer Dubins path from $o^{s,k}_j$ to $o^{g,l}_j$ is infeasible; bottom: transitioning to a new pushing pose when the object is transferred to a prerelocation enables a feasible push transfer. (\ref{['fig:prerelob']}) Zoom in on a prerelocation. A robot path $\mathcal{P}_1$, from $p_j^{s,k}$ to $p_j^{pre,a}$ (shown in green) transfers the object from $o_j^{s,k}$ to $o_j^{pre}$. After the robot transits to a new pushing pose $p_j^{pre,b}$ (transit path not shown), $\mathcal{P}_2$ transfers the object to $o_j^{g,l}$.
• Figure 5: For a car-like robot, a Dubins path describes the optimal path of transitioning from a start pose $p^s$ to a goal pose $p^g$. A Dubins path can be of two different type families, i.e., $CSC$ or $CCC$, where $S$ is a a straight line segment, and $C\in \{L, R\}$ denotes a left or right arc. When the distance $d$ is too small relative to the respective heading difference, only $CCC$ paths are admissible (the short-distance regime), yielding longer paths with a wider spatial sweep shkel2001classification. Circles indicate minimum–turning-radius arcs; $\alpha$ and $\beta$ are the start and goal headings.

Theorems & Definitions (2)

• Theorem IV.1
• proof