Conquer:
Continual Quadruped Robots Coordination via Semantic Skill Discovery

Anonymous Authors
Simulation and Real-Robot Demonstration
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 panel
Detailed task configurations and initial state distributions for the 14-task benchmark, covering 1D/2D/3D object transport across flat and rough terrain with variable-size robot teams.
Simulation task topview
Top-down view of all 14 simulation tasks showing the full trajectory and interaction patterns for each coordination scenario.

Main Results

Main results table
Comparison of Conquer against state-of-the-art baselines across the 14-task benchmark. Metrics include average success rate, forward transfer, and forgetting, evaluated over five random seeds.
Training growth curves
Per-task success rate growth curves across methods. Conquer rapidly acquires new skills while preserving performance on previously learned tasks, showing effective catastrophic forgetting prevention across the full task stream.

Case Study

t-SNE visualization of task semantics
t-SNE visualization of learned task semantics. Tasks with similar coordination patterns form meaningful clusters, validating the semantic descriptor's ability to capture task structure.
Leave-one-out analysis
Leave-one-out ablation of the skill library components. Each component contributes to overall coordination performance, with the semantic retrieval module providing the largest marginal gain.

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.
1-dog command waveform
1-Dog Tracking
2-dog command waveform
2-Dog Coordination
3-dog command waveform
3-Dog Coordination
4-dog command waveform
4-Dog Coordination

Command tracking waveforms for 1–4 dog deployments. Each subplot shows the commanded (dashed) versus actual (solid) velocity profiles, demonstrating accurate real-world command following across all team sizes.

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 enhanced payload capacity, broader contact coverage, and improved adaptability to challenging tasks. Existing methods for multi-quadruped manipulation typically focus on predefined or closed task families, often relying on multi-agent reinforcement learning (MARL) to train task-specific coordination policies. However, such methods struggle in open-ended continual learning settings, where tasks arrive sequentially and robots are expected to acquire new coordination skills while reusing previously learned ones without catastrophic forgetting.

To address this challenge, we propose Conquer, a semantic skill-library framework that formulates continual multi-quadruped coordination as a retrieve-adapt-update process. First, to accommodate varying team sizes across tasks, we design a team-structured Self-Allies-Goal (SAG) backbone that supports variable-cardinality robot teams by explicitly modeling each robot's own state, teammate context, and task goal. For each incoming task, Conquer constructs a task-level semantic descriptor from pre-execution information and retrieves a relevant skill from the library for adaptation. After successful execution, Conquer updates the skill library by extracting trajectory-level semantic descriptors and organizing them according to semantic distance, thereby enabling continual skill accumulation and cross-task knowledge transfer.

Simulation experiments show that Conquer achieves a final average success rate of 95.6%, demonstrating strong forward transfer and negligible catastrophic forgetting. Real-world rollouts on Unitree Go2 teams further validate the deployment feasibility of Conquer for practical multi-quadruped coordination.

Team

Anonymous Team

BibTeX

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