Toward Reliable Sim-to-Real Predictability for MoE-based Robust Quadrupedal Locomotion

TL;DR

We address the sim-to-real transfer challenge in proprioception-only quadrupedal locomotion by coupling a Mixture-of-Experts latent representation with RoboGauge, a predictive, sim-to-sim transfer framework. The MoE policy is trained within a Concurrent Teacher-Student regime and evaluated across diverse terrains, while RoboGauge provides six proprioception-based metrics across seven terrains with domain randomization to quantify transferability and guide policy selection. Physical deployment on a Unitree Go2 demonstrates robust performance on unseen terrains, high-speed capability up to 4 m/s, and an emergent narrow-width gait, outperforming baseline methods. This framework reduces hardware risk, accelerates deployment, and offers a scalable path toward broader morphologies and perception-enhanced control.

Abstract

Reinforcement learning has shown strong promise for quadrupedal agile locomotion, even with proprioception-only sensing. In practice, however, sim-to-real gap and reward overfitting in complex terrains can produce policies that fail to transfer, while physical validation remains risky and inefficient. To address these challenges, we introduce a unified framework encompassing a Mixture-of-Experts (MoE) locomotion policy for robust multi-terrain representation with RoboGauge, a predictive assessment suite that quantifies sim-to-real transferability. The MoE policy employs a gated set of specialist experts to decompose latent terrain and command modeling, achieving superior deployment robustness and generalization via proprioception alone. RoboGauge further provides multi-dimensional proprioception-based metrics via sim-to-sim tests over terrains, difficulty levels, and domain randomizations, enabling reliable MoE policy selection without extensive physical trials. Experiments on a Unitree Go2 demonstrate robust locomotion on unseen challenging terrains, including snow, sand, stairs, slopes, and 30 cm obstacles. In dedicated high-speed tests, the robot reaches 4 m/s and exhibits an emergent narrow-width gait associated with improved stability at high velocity.

Figures (18)

• Figure 1: Our proposed framework integrates a Mixture-of-Experts architecture for terrain and command representation with the RoboGauge assessment suite to quantify sim-to-real transferability through sim-to-sim metrics. This closed-loop design enables reliable policy selection to facilitate robust deployment for agile locomotion across diverse challenging environments based solely on proprioception.
• Figure 2: Comparative analysis against one-stage proprioceptive methods including CTS, HIM, and DreamWaQ. Within the RoboGauge framework, each axis reflects average performance on a specific terrain and serves as a reliable proxy to quantify sim-to-real capability. Our architecture consistently outperforms or matches previous state-of-the-art across all evaluated terrains under RoboGauge's metrics.
• Figure 3: The RoboGauge evaluation architecture consists of three hierarchical stages. (A) Base Pipeline serves as a single evaluation environment by incorporating specific terrains and domain randomization. (B) Multi/Level Pipeline highlights the Multi/Level Pipeline where parallel evaluations across diverse random seeds. (C) Stress Pipeline depicts the Stress Pipeline which triggers comprehensive testing across the entire terrain suite to synthesize the final score.
• Figure 4: Comparison of RoboGauge scores and terrain level curves across various baselines during training. Stable RoboGauge scores despite fluctuating terrain levels demonstrate that training levels fails to accurately represent model performance.
• Figure 5: Comparison of maximum terrain levels across varying friction coefficients as evaluated by RoboGauge.