Domain Randomization for Robust Sampling-Based Model Predictive Control on Bipedal Wheeled Robots
model-predictive-control, domain-randomization, bipedal-wheeled-robots, world-models, 2026
Reinforcement learning with domain randomization achieves robust sim-to-real transfer for legged robots but produces fixed policies lacking online adaptability. GPU-accelerated sampling-based MPC offers online adaptability and interpretable objectives but suffers from brittle sim-to-real transfer. This work proposes integrating domain randomization directly into the sampling-based MPC loop (DR-MPC) for robust sim-to-real transfer on the LIMX Tron1 bipedal wheeled robot, without requiring online parameter estimation or per-robot tuning. Two complementary instantiations are presented: (1) DR-MPC with physics simulation, which randomizes dynamics parameters across GPU-parallelized rollouts and selects actions via MPPI/CEM with tunable risk-sensitive cost aggregation; and (2) DR-MPC with learned surrogate dynamics, which trains a neural dynamics model conditioned on physical parameters on domain-randomized data. Key design choices including parameter selection, randomization range, cost aggregation, and sample budget allocation are ablated on velocity tracking, push recovery, slope traversal, and turning tasks.
