Estimation and Control of Motor Core Temperature with Online Learning of Thermal Model Parameters: Application to Musculoskeletal Humanoids

TL;DR

The paper addresses motor core temperature management for robots with strict weight/shape constraints by online learning of a two-resistor thermal model to accurately estimate $c_{1}$ from measurable $c_{2}$ and actuation $f$. It introduces parameter updates for $P_{1}$–$P_{5}$, anomaly detection via parameter drift, and a backpropagation-based thermal controller that computes a sequence $f^{limit}$ to bound $c_{1}$ while enabling continuous motion, including a muscle-length limiter for safe actuation. The approach is validated in simulation, with a real actuator, and on the Musashi musculoskeletal humanoid, demonstrating real-time feasibility and enabling high-tidelity temperature management across multiple actuators. This software-based thermal management enables safe, sustained humanoid motion without heavy hardware cooling, and the method’s interpretability via explicit thermal-parameter changes supports rapid anomaly diagnosis and maintenance.

Abstract

The estimation and management of motor temperature are important for the continuous movements of robots. In this study, we propose an online learning method of thermal model parameters of motors for an accurate estimation of motor core temperature. Also, we propose a management method of motor core temperature using the updated model and anomaly detection method of motors. Finally, we apply this method to the muscles of the musculoskeletal humanoid and verify the ability of continuous movements.

Figures (14)

• Figure 1: System overview.
• Figure 2: Overview of a basic two-resistor thermal model.
• Figure 3: Difference of transition of $c_{2}$ when applying $f=200$ [N] between new and old motors
• Figure 4: Muscle actuator configuration in Musashi: sensor-driver integrated muscle module asano2015sensordriver and miniature bone-muscle module kawaharazuka2017forearm.
• Figure 5: Transition of errors in thermal model parameters between simulator and current model during online learning.