Research Analyzer
← Back ICRA 2026

Heterogeneous Skill Learning for Asynchronous Multi-Robot Relay Pushing in Complex Environments

Hui Zhi, David Navarro-Alarcon

PDF

AI summary

Key figure (auto-extracted from paper)
Combining curriculum learning and affordance-guided reinforcement learning enables reliable, sensor-free multi-robot relay pushing in complex environments.
Multi-robot collaboration nonprehensile manipulation relay pushing deep reinforcement learning affordance learning curriculum learning

Problem

Multi-robot systems struggle with coordination and real-time control when performing long-horizon pushing tasks in cluttered or constrained spaces, while existing methods often rely on restrictive trajectory constraints or computationally heavy optimization.

Approach

The framework decomposes relay pushing into three learned skills using Soft Actor-Critic, applying curriculum training for open spaces and an affordance network to guide feasible contact points in narrow corridors without external force sensors.

Key results

  • Geometry-aware pushing enables contact-rich manipulation without force sensors
  • Affordance-guided policy achieves ~0.8 success rate in constrained corridors, significantly outperforming baselines
  • Curriculum training ensures stable convergence (0.8–0.9 success) across progressively challenging room pushing tasks
  • Learned skills form a reusable primitive library for asynchronous multi-robot relay transport

Why it matters

Provides a scalable, sensor-light coordination strategy for cooperative multi-robot manipulation in cluttered or narrow real-world environments.

Abstract

This paper presents a heterogeneous skill learning frame- work for asynchronous multi robot relay pushing in complex and cluttered environments. To support cooperative relay transportation, we construct a skill library comprising room robot pushing, corridor helper pushing, and standby behaviors. We further propose a geometry aware pushing strategy that enables contact rich manipulation without relying on external force sensors. For the room robot, curriculum learning is adopted to decompose training into an approach to parcel phase and a parcel to target pushing phase, thereby improving training stability and task progression. For long horizon transportation in con- strained corridors, an affordance network is introduced to model the local feasibility of pushing actions, providing structured guidance that improves policy learning efficiency. The overall framework combines Soft Actor Critic with Dijkstra based reachability maps to coordinate the“Corridor Helper Pushing” skills. Experimental results demonstrate high success rates across progressive curriculum lessons, suggesting that the proposed framework provides an effective skill primitive for cooperative multi robot transportation.

Index terms

Mobile Manipulation Deep Learning in Grasping and Manipulation Multi-Robot Systems

Related papers