Why Do Quadruped Locomotion Policies Fail When Transferred from Simulation to Real Hardware? A Systematic Analysis of Failure Modes Across MPC, RL, and IL
Amr Ahmed
12/08/2026
Sim-to-real transfer remains one of the main challenges in deploying quadruped controllers, as discrepancies between simulated and real-world dynamics frequently cause policies that succeed in simulation to fail on hardware. This paper presents a structured, systematic review comparing how three dominant control paradigms: Model Predictive Control, Reinforcement Learning, and Imitation Learning handle sim-to-real transfer in quadruped locomotion, alongside hybrid approaches that combine elements of all three. Reviewed studies were organized using a four-dimensional taxonomy: control paradigm, gap source, failure manifestation, and mitigation strategy. Results show that MPC offers strong physical interpretability and constraint handling but is vulnerable to inaccurate contact and inertia models. RL demonstrates robust terrain adaptation but is prone to overfitting to simulation physics. IL produces structured, natural-looking behaviours but struggles when simulation observations do not match real-world sensing constraints. Hybrid architectures, which embed learned components into model-based frameworks or combine multiple paradigms, showed the most reliable transfer by compensating for the weaknesses of each paradigm. The study concludes that standardized evaluation benchmarks and unified sim-to-real assessment pipelines are needed to enable rigorous cross-paradigm comparison in future work.