Conquer:
Continual Quadruped Robots Coordination via Sementic Skill Discover

Anonymous Authors
Real Robot Demonstrations
Demo Videos Abstract BibTeX

Method Overview

Conquer method overview
Overview of the Conquer framework. A semantic skill-library architecture that models continual multi-quadruped coordination as a retrieve-adapt-update process built on a team-structured Self-Allies-Goal (SAG) backbone.

Simulation Tasks

Simulation task overview
The 14-task Isaac Lab benchmark spans diverse coordination scenarios including object transport, formation control, and cooperative manipulation across variable-size robot teams.

Comparison with Baselines

Training growth curves
Per-task success rate growth curves. Conquer rapidly acquires new skills while preserving performance on previously learned tasks, showing effective catastrophic forgetting prevention across the full task stream.

Case Study

Case study analysis
Semantic transfer analysis showing how skills learned in earlier tasks accelerate adaptation to new coordination scenarios. t-SNE visualization reveals meaningful clustering of task semantics.

Real-Robot Deployment

Real robot deployment
Real-world deployment on Unitree Go2 teams validates simulation-to-real transfer. The framework successfully scales from single-robot to multi-robot coordination on physical hardware.

Real-robot Demonstrations

Single-dog Tracking

Single-dog Fixed-point

Two-dog Fixed-point

Three-dog Tracking

Three-dog Fixed-point

Four-dog Fixed-point

Simulation Demonstrations

1D Cube (Flat)

1D Cube (Rough)

1D Cuboid (Flat)

1D Cuboid (Rough)

1D Triangle (Flat)

1D Triangle (Rough)

2D Cuboid (Flat)

2D Cuboid (Rough)

2D Triangle (Flat)

2D Triangle (Rough)

3D Cuboid (Flat)

3D Cuboid (Rough)

3D Triangle (Flat)

3D Triangle (Rough)

Abstract

Multi-quadruped coordination has attracted increasing attention due to its stronger payload capacity, broader contact coverage, and better adaptability to challenging tasks. However, existing multi-quadruped manipulation methods typically focus on a closed task family, and thus struggle to continually acquire and reuse coordination skills in open-ended task streams.

To address this issue, we propose Conquer, a semantic skill-library framework that models continual multi-quadruped coordination as a skill retrieve-adapt-update process. For each incoming task, Conquer first builds a semantic descriptor from pre-execution information and retrieves a relevant skill from the library; it then either reuses the retrieved policy or adapts a new skill adapter through MAPPO, and finally updates or expands the library using descriptors extracted from successful trajectories. This process is built on a team-structured Self-Allies-Goal (SAG) backbone that supports variable-size robot teams while keeping stored skills parameter-isolated.

Experiments on a 14-task Isaac Lab benchmark show that Conquer achieves 0.9626 final average success rate with strong forward transfer and no forgetting. Real-world rollouts on Unitree Go2 teams further demonstrate its deployment feasibility for real-world multi-quadruped coordination.

Team

Anonymous Team

BibTeX

@misc{conquer2026,
  title  = {Conquer: Continual Quadruped Robots Coordination via Sementic Skill Discover},
  author = {Anonymous},
  year   = {2026},
}