A Study of Integrating Thermal Models into Actuators in Physics Simulators for High-fidelity Sim-to-real Transfer

The performance gap between simulated and real-world robots, known as the Sim-to-real gap, remains a critical challenge in robotics. While methods such as system identification and domain randomization address time-invariant mechanical parameters, they often overlook the significant performance degradation caused by actuator overheating during prolonged operation. High-current usage in modern QDD (Quasi-Direct Drive) actuators generates Joule heating, reducing motor torque constant and exacerbating the Sim-to-real gap. This study proposes a method to bridge this gap by integrating an LPTN (Lumped Parameter Thermal Network) model directly into the Isaac Gym physics simulation loop. The model estimates the real-time rotor temperature, computes torque constant degradation, and applies the resulting torque loss within the simulator's physics engine. The proposed simulation with thermal network model is validated against a baseline simulation without it and a real 1-DoF robotic testbed under identical long-duration sinusoidal inputs. Experimental results demonstrate that the proposed simulation incorporating the thermal model accurately reproduces the performance degradation observed in the real system with a temperature RMSE of less than 3.79 °C and a 29.1% relative error in final amplitude decay, which the baseline simulation fails to capture.