Continual Quadruped Robots Coordination via Semantic Skill Discovery

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. Simulation and real-robot demonstration videos are available at: https://conquer-project.pages.dev/.

• The paper presents a semantic skill library with a retrieve-adapt-update process that enables continual learning across variable robot teams.
• It uses the SAG backbone and LoRA-based adapters to achieve a 95.6% success rate and strong forward transfer while preventing catastrophic forgetting.
• Experiments in simulation and real-world deployments validate its scalability, robust task performance, and effective skill transfer.

Continual Quadruped Coordination via Semantic Skill Discovery: An Expert Overview

Motivation and Problem Statement

The coordination of multiple quadruped robots presents a pivotal route for extending manipulation payload, contact coverage, and task adaptability beyond the limits of single platforms. Prior research has predominantly relied on MARL approaches tailored to specific, static task regimes, often deploying fixed-size teams and rigid task boundaries. This creates fundamental barriers for open-ended, continual learning scenarios where task distributions shift, team sizes vary, and catastrophic forgetting must be avoided. The challenge is to engineer mechanisms that enable sequential task ingestion, skill reuse, and robust cross-task transfer, without loss of previously acquired competencies and without explicit re-training or task boundary specification.

Algorithmic Framework

Conquer introduces a semantic skill-library approach with a retrieve-adapt-update workflow tightly integrating semantic information into skill discovery and transfer. The core components are:

• SAG Backbone: The Self-Allies-Goal (SAG) architecture provides a unified policy interface consistent across variable-cardinality robot teams. Each agent observation is decomposed into self, ally, and goal tokens, with cross-attention mechanisms fusing contextual information and yielding team-aware, task-adaptive latent representations.
• Skill Library: Each coordination skill is parameterized via LoRA-based lightweight adapters (LoRA, LocHead), linked to semantic descriptors. Skills are indexed by mean embeddings generated from trajectory rollouts and task instructions, facilitating explicit semantic retrieval.
• Semantic Descriptor Interface: A VLM-to-embedding pipeline first maps pre-execution task instructions and visual observations into a shared embedding space. This space is used to compute nearest-neighbor matches between incoming task queries and skill library entries, using Euclidean distance as a measure of semantic compatibility.
• Retrieve-Adapt-Update Loop: Each incoming task initiates retrieval of the closest semantic skill, adaptation via a frozen SAG backbone and MAPPO optimization, and library update contingent on semantic descriptor overlap—either replacing the existing entry or adding a new skill if semantic distance exceeds a threshold.

Experimental Results

Simulation Benchmark

On a 14-task Isaac Lab push-to-goal stream covering variable robot team sizes, object geometries, and terrain, Conquer demonstrates:

• Final SR: 95.6% average success rate, exceeding both the multitask joint-training baseline (93.5%) and all continual learning baselines.
• Forward Transfer (FWT): 11.3%, the highest observed among sequential online methods, validating the effectiveness of semantic initialization.
• Backward Transfer (BWT): 0.0%, indicating negligible catastrophic forgetting over the task stream.

Ablation studies highlight the necessity of semantic retrieval (Random selection reduces SR by 22.2%), LoRA adaptation (Multihead-only reduces SR by 8.8%), and initialization from prior skills (Scratch reduces SR by 6.4%). Semantic retrieval systematically improves zero-shot skill matching; the top-3 semantic neighbors yield substantially higher transfer rates compared to non-neighbors, although optimality is not always guaranteed by semantic similarity alone.

Real-World Deployment

Policy transfer from simulation to Unitree Go2 teams is achieved via hierarchical control stacks (high-level velocity policies, low-level locomotion controllers), with adaptive domain randomization addressing real-world physics mismatches. Physical deployments on tasks with 1–4 robots show successful object-to-goal manipulation, final goal distances reduced to sub-0.3m with dynamic adaptation to differential payloads and terrain. The semantic skill-library architecture exhibits robust transferability and real-time feasibility in practical robotic systems.

Implications and Future Directions

Conquer marks an overview of parameter isolation, semantic retrieval, and robust continual learning within the paradigm of variable-team multi-quadruped coordination. The demonstrated semantic interface enables scalable and flexible skill management, offering a principled approach to knowledge transfer and retention across open-ended task streams.

Practical implications are notable for industrial logistics, search-and-rescue, and adaptive automation scenarios where heterogeneous robot teams must continually update and reuse manipulation skills. The framework suggests that semantic embeddings, when leveraged judiciously, allow for efficient allocation and reuse of skills, reducing the need for exhaustive retraining and dataset replay.

Theoretical implications open the avenue for incorporating closed-loop skill management, wherein physical transfer statistics and training dynamics could further inform semantic retrieval. Future research may expand towards compositional skill frameworks, heterogeneous agents, and fine-grained semantic decomposition to optimize knowledge transfer across broader loco-manipulation domains.

Conclusion

Conquer provides a semantic skill-library scaffold for continual multi-quadruped coordination, delivering strong numerical results in task success, forward transfer, and retention under realistic robot deployments. Its retrieve-adapt-update workflow with SAG backbone and parameter-isolated skill adapters establishes a robust foundation for future, semantically-guided, continual learning in multi-agent robot systems. Limitations remain in the reliance on static semantic descriptors and the lack of explicit modeling for physical transferability; addressing these could further generalize the approach to complex and heterogeneous environments.

(2606.08102)