Cycloidal Quasi-Direct Drive Actuator Designs with Learning-based Torque Estimation for Legged Robotics

Alvin Zhu*, Yusuke Tanaka*, Fadi Rafeedi, and Dennis Hong

* denotes co-first authorship

Presented at ICRA 2025

Paper

Code

Video

Bottom graph plots the estimated torque vs. sensor-less predicted torque from our framework

Overview

This research paper introduces a Cycloidal Quasi-Direct Drive (C-QDD) actuator designed for legged robotics, emphasizing its compact, high-torque capabilities and mechanical robustness. The innovative design leverages cycloidal gears, which outperform traditional planetary gears in torque density, low backlash, and resistance to dynamic loads. To address challenges like torque ripple and the sim-to-real gap caused by the cycloidal mechanism's nonlinear dynamics, the paper integrates a Gated Recurrent Unit (GRU)-based torque estimation framework. This framework models actuator dynamics with an almost 99% accuracy, enhancing reinforcement learning performance and adaptability for legged robots. The C-QDD's and torque estimation framework's performance is evaluated through hardware benchmarks, showcasing its efficiency in high-dynamic environments and its potential for advancing agile robotics.

Overview of the Torque Estimation Framework for online prediction.

The torque estimation framework uses a Gated Recurrent Unit (GRU)-based architecture to model the nonlinear dynamics of the C-QDD actuator. By processing historical joint states, including position error, velocity, and optionally acceleration, it accurately predicts high-frequency torque ripple and nonlinearities. The framework's PVA-GRU variant excels in precision, leveraging acceleration data for better performance in dynamic scenarios. This approach reduces the sim-to-real gap, enabling accurate torque control and enhancing the integration of C-QDD actuators into reinforcement learning for legged robotics.

The C-QDD actuator combines a compact 10:1 cycloidal gearbox with a quasi-direct drive system, optimized for high torque and minimal backlash in legged robotics. Its design features precision-machined cycloidal disks, eccentric cams, and counterbalancing disks to reduce torque ripple and ensure smooth operation under dynamic loads. Built with high-strength 4140 alloy steel and manufactured with tight tolerances, the C-QDD achieves superior load distribution, efficiency, and durability, making it ideal for demanding robotic applications.

An exploded view of the C-QDD Design

Citation

If you use this work or find it helpful, please consider citing: (bibtex)

@misc{zhu2024cycloidalquasidirectdriveactuator,
title={Cycloidal Quasi-Direct Drive Actuator Designs with Learning-based Torque Estimation for Legged Robotics},
author={Alvin Zhu and Yusuke Tanaka and Fadi Rafeedi and Dennis Hong},
year={2024},
eprint={2410.16591},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2410.16591},
}