Fully automated system identification for sim-to-real transfer

Develop a fully automated system identification methodology for legged robot platforms used in ZEST (Boston Dynamics Atlas, Unitree G1, and Boston Dynamics Spot) that estimates dynamics and actuator parameters directly from data to reduce sim-to-real mismatch in policies trained in simulation, particularly for complex phenomena that are not adequately captured by first-principles models.

Background

ZEST relies on simulation-trained policies that are deployed zero-shot to hardware, which makes the accuracy of the simulation models and low-level controller gains crucial for successful transfer. The paper introduces practical modeling procedures, including parallel-linkage actuator approximations and armature-dependent PD gain selection, and demonstrates robust sim-to-real performance across Atlas, G1, and Spot.

Despite these advances, the authors explicitly flag the absence of a fully automated system identification pipeline as an outstanding challenge. They note that complex real-world phenomena are difficult to capture with first-principles models and likely require data-driven approaches, motivating the need for automated identification to further tighten the sim-to-real gap.