ALORE: Autonomous Large-Object Rearrangement with a Legged Manipulator

TL;DR

This work introduces ALORE, a complete autonomous system for rearranging large objects with a legged manipulator. It combines a hierarchical reinforcement learning framework with a novel Interaction Configuration Representation based on graphs and a learning-based object velocity estimator, all orchestrated by a task and motion planning module to enable multi-object, long-horizon rearrangement. The approach demonstrates robust performance in simulation and real-world experiments, including long-distance tasks over corridors and large indoor spaces, rearranging dozens of chairs with a single policy and minimal failures. Open-source releases accompany the system to accelerate public adoption and further research in loco-manipulation and large-object rearrangement. The results indicate significant advances in generalization across object types, safety during interaction, and the integration of perception, planning, and control for autonomous large-object rearrangement.

Abstract

Endowing robots with the ability to rearrange various large and heavy objects, such as furniture, can substantially alleviate human workload. However, this task is extremely challenging due to the need to interact with diverse objects and efficiently rearrange multiple objects in complex environments while ensuring collision-free loco-manipulation. In this work, we present ALORE, an autonomous large-object rearrangement system for a legged manipulator that can rearrange various large objects across diverse scenarios. The proposed system is characterized by three main features: (i) a hierarchical reinforcement learning training pipeline for multi-object environment learning, where a high-level object velocity controller is trained on top of a low-level whole-body controller to achieve efficient and stable joint learning across multiple objects; (ii) two key modules, a unified interaction configuration representation and an object velocity estimator, that allow a single policy to regulate planar velocity of diverse objects accurately; and (iii) a task-and-motion planning framework that jointly optimizes object visitation order and object-to-target assignment, improving task efficiency while enabling online replanning. Comparisons against strong baselines show consistent superiority in policy generalization, object-velocity tracking accuracy, and multi-object rearrangement efficiency. Key modules are systematically evaluated, and extensive simulations and real-world experiments are conducted to validate the robustness and effectiveness of the entire system, which successfully completes 8 continuous loops to rearrange 32 chairs over nearly 40 minutes without a single failure, and executes long-distance autonomous rearrangement over an approximately 40 m route. The open-source packages are available at https://zhihaibi.github.io/Alore/.

Figures (13)

• Figure 1: Illustration of a large-object rearrangement system with a legged manipulator. (a) Workflow of the proposed system, including tailored perception, robot-object system planning, and a hierarchical RL-based controller. (b) Training pipeline of the hierarchical controller, which consists of the low-level WBC and the high-level object controller. After training, the learned policy can be directly deployed on a legged manipulator without further fine-tuning.
• Figure 2: Illustration of interaction configuration representation. (a) Interaction configuration examples when the robot interacts with three different types of objects. The red line segments represent the topology of the interaction configuration, while the pink, blue, and green nodes correspond to the nodes on the robot, the manipulator, and the objects, respectively. (b) The robot-object interactive configuration is modeled as an IG, where each node or edge contains the features of the robot and object. The graph is then processed by a GNN to extract the graph-level feature. (c) Details of the message generation and aggregation process during GNN training.
• Figure 3: System setup of the legged manipulator.
• Figure 4: Comparison of training rewards over 6000 episodes against three baselines and two ablation variants.
• Figure 5: Comparison of velocity tracking curves of bin (a), chair (b), and table (c) under five kinds of methods.