Implicit Behavior Coordination from Unlabeled Sub-Task Demonstrations for Rearrangement Tasks

Jul 16, 2026

Long-horizon robotic rearrangement is commonly modeled as skill sequencing, requiring predefined skills, labeled demonstrations or boundaries, and switching logic. This engineering overhead grows rapidly with more behaviors and longer horizons. The approach instead learns rearrangement through implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are extracted from mixed data and coordinated via value-guided action selection. Habitat experiments validate improved performance, critic-guided candidate selection importance, and strong scaling to larger repertoires and extended horizons.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

Related Work

arXiv :2607 .09234v1 [ cs .RO] 10 Jul 2026 Implicit-Behavior Coordination from Unlabeled Sub-Task Demonstrations for Rearrangement Tasks Ahmed Shokry1,2 Usama Ahmed Siddiquie1 Sicong Pan1,2 Maren Bennewitz1,2 1Humanoid Robots Lab and Center for Robotics, University of Bonn, Germany 2Lamarr Institute for Machine Learning and Artificial Intelligence, Bonn, Germany Abstract: Long-horizon robotic rearrangement tasks are often treated as skill sequencing problems, requiring predefined skills, skill labels, or boundaries, and task-specific switching logic. Although effective, such explicit skill abstractions can become difficult to scale as the number of behaviors and the task horizon increase. We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection. Experiments in Habitat rearrangement tasks support this formulation in three ways. First, our method outperforms task-specific imitation baselines on more complex rearrangement tasks and approaches an oracle-planner baseline with behaviorcloned skills, while using no oracle task plan or skill-labeled full-task demonstrations. Second, ablations show that reliable critic-guided candidate selection is essential for coordinating multi-modal behaviors. Third, scaling experiments show that the method handles larger behavior repertoires and maintains stronger performance than task-specific imitation baselines as chained targets extend the horizon. These results suggest that explicit skill abstraction is not a prerequisite for long-horizon rearrangement, and that implicit-behavior coordination offers a promising data-driven alternative to explicit skill-based pipelines. Keywords: Imitation Learning 1 Introduction Long-horizon robotic rearrangement is often formulated as a skill sequencing problem [1, 2] . A robot is assumed to possess a set of predefined skills, such as navigation, object interaction, picking, and placing, and the main challenge is to decide which skill to execute at each stage of the task. This formulation has led to effective systems based on modular skill libraries, task planners, highlevel skill policies, or language-conditioned coordinators [3, 4, 5] . However, it also introducesa substantial engineering burden: skills must be manually defined, demonstrations must often be annotated with skill labels or boundaries, interfaces between skills must be specified, and switching logic must be designed or learned for each task family. As the number of skills grows or the task horizon becomes longer, maintaining this explicit coordination layer becomes increasingly difficult. In this work, we investigate whether explicit skill abstraction is a prerequisite for long-horizon rearrangement. Rather than defining a skill library and learning a coordinator over discrete skill identities, we formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations. The demonstrations contain skill-like behaviors, but these behaviors are not exposed to the learning algorithm as explicit skill labels, skill-specific policies, or hand-designed transition rules. Our goal is to learn directly from this mixed unlabeled behavior data and coordinate the learned behaviors according to the current state and the desired final state. Figure 1: Overview of implicit-behavior coordination. The Flow Matching policy samples multiple candidate action chunks from a learned multi-modal distribution conditioned on the current observation context. These candidates may correspond to different skill-like behaviors, but no skill identity or behavior label is provided. The critic scores each candidate according to the task objective and the action chunk with the highest expected return is chosen for execution in the environment. Repeating this procedure across task execution enables our fra

FAQs

What evidence supports the effectiveness and scalability of the method?

Experiments in Habitat show improved results over task-specific imitation baselines and near-oracle performance characteristics, ablations confirm the need for critic-guided candidate selection, and scaling experiments maintain stronger performance as behavior repertoires and chained-target horizons increase.

Jul 16, 2026

Upload by Seraphina