Constraint-Aware Diffusion Priors for High-Fidelity and Versatile Quadruped Locomotion

2026-05-09

AI Summary

This work addresses key limitations in existing quadrupedal locomotion control methods—namely mode collapse induced by GAN discriminators, heading drift during complex maneuvers, and hardware-unsafe behaviors arising from unenforced actuator constraints. To overcome these challenges, the authors propose Diff-CAST, a novel framework that introduces diffusion models for the first time into quadruped motion prior modeling, effectively mitigating mode collapse by replacing conventional GAN discriminators. The approach further incorporates a symmetry-augmented command-conditioning mechanism to eliminate heading drift and integrates constrained reinforcement learning to guarantee actuator safety. Combined with Sim2Real transfer and stylized imitation learning, Diff-CAST enables high-fidelity, multimodal motion generation and seamless transitions on real robots, achieving both drift-free trajectory tracking and strict compliance with hardware constraints.

Abstract

Reinforcement learning combined with imitation learning has significantly advanced biomimetic quadrupedal locomotion. However, scaling these frameworks to massive, multi-source datasets exposes fundamental bottlenecks. First, traditional GAN-based discriminators are prone to mode collapse, struggling to capture diverse motion distributions from uncurated datasets. Second, existing kinematic priors suffer from out-of-distribution (OOD) tracking conflicts, leading to severe unintended heading drifts during complex maneuvers. Furthermore, deploying unconstrained priors to physical hardware poses critical safety risks by disregarding actuator dynamics. To overcome these challenges, we propose Diff-CAST (Diffusion-guided Constraint-Aware Symmetric Tracking), a novel motion prior framework leveraging the multi-modal distribution modeling capabilities of diffusion models for stylistic rewards. Diff-CAST effectively replaces traditional GAN discriminators, unlocking robust data scaling on heterogeneous collections. To ensure high-fidelity intent execution and reliable real-world deployment, we introduce a comprehensive Sim2Re architecture integrating Symmetric Augmented Command Conditioning (SACC) for drift-free tracking, and Constrained RL for hardware safety. Experiments on a quadruped demonstrate that Diff-CAST mitigates mode collapse, enables seamless transitions between diverse skills, and ensures robust, hardware-compliant locomotion.

Problem

• quadruped locomotion
• mode collapse
• out-of-distribution tracking
• actuator dynamics
• motion priors

Innovation

• diffusion models
• constraint-aware locomotion
• mode collapse mitigation
• Sim2Re
• symmetric command conditioning