TL;DR
This paper introduces a novel bio-inspired robotic shoulder joint that achieves precise control using self-sensing artificial muscles, eliminating the need for external sensors. This innovation leads to a more compact, compliant, and cost-effective design, paving the way for advanced robotic systems that mimic biological movement.
What Problem Does This Paper Address?
Current robotic shoulder designs often struggle to replicate the natural flexibility, compactness, and efficiency of their biological counterparts. A significant hurdle is the reliance on external sensors like rotary encoders. These sensors add bulk, restrict mechanical design, and increase complexity and cost. The paper aims to overcome this by developing a robotic shoulder that can sense its own position and movement internally, much like biological muscles, thereby enabling more streamlined and capable robotic joints.
Methods
The core of this work lies in the development of a two-degree-of-freedom (DoF) robotic shoulder joint powered by antagonistic pairs of hydraulically amplified self-healing electrostatic (HASEL) actuators. The key innovation is the integration of a self-sensing capability directly into these artificial muscles. This is achieved by decoupling the high-voltage actuation electrodes from a separate pair of low-voltage sensing electrodes. By applying a low-voltage sinusoidal signal across these sensing electrodes, the system can measure the voltage drop, which changes proportionally to the deformation of the HASEL actuator. This voltage signal acts as a proprioceptive feedback, allowing the system to estimate the actuator's displacement and, consequently, the joint's position without external sensors.
The system employs a PID controller for actuation. A crucial aspect is the mapping between the task space (desired end-effector position) and the tendon space (required HASEL displacements), and then from the self-sensing voltages to the estimated task space position. This mapping is learned using a third-order polynomial regression, trained with ground truth data from a motion capture system. The control architecture supports two configurations: one using the self-sensing feedback and another using motion capture data as a benchmark.
Datasets
The paper does not explicitly mention the creation or use of distinct datasets for training. Instead, it describes a process of system identification and model training using data acquired during experimental trials. Specifically, ground truth data from a Qualisys motion capture system (tracking the manipulator's six DoF at 150 Hz) was used to train the self-sensing voltage to displacement estimation model. This data served as the benchmark for evaluating the self-sensing control performance.
Key Insights and Ideas
- Self-Sensing Artificial Muscles: The most significant insight is the successful implementation of self-sensing capabilities within HASEL actuators. This eliminates the need for external sensors, a major step towards more compact and bio-inspired robotic designs.
- Bio-Inspired Design: The robotic shoulder's design draws direct inspiration from the human shoulder's glenohumeral joint and antagonistic muscle structure, aiming to replicate its dexterity and efficiency.
- Decoupling Actuation and Sensing: The clever separation of high-voltage actuation electrodes from low-voltage sensing electrodes is the key to achieving reliable self-sensing without interfering with the primary actuation function.
- Task Space Control without External Sensors: The paper demonstrates that closed-loop control in the task space is achievable solely through the inherent self-sensing capabilities of the actuators, a significant advancement for robotic joint design.
- Tendon-Based Force Transmission: Utilizing tendons for force transmission from the HASEL actuators to the output link is an effective strategy to minimize frictional losses and enable quick movements.
Key Results
The paper demonstrates closed-loop control of a two-DoF robotic shoulder using self-sensing feedback. When compared to a benchmark using motion capture data, the self-sensing control achieved an RMSE of 4.245 mm for a lemniscate trajectory (a 48.0% increase over the benchmark) and 3.407 mm for a star trajectory (a 21.8% increase). While there is an increase in error compared to the benchmark, these results show promising performance and a step towards sensorless control. The system also demonstrated a deflection of more than 80° in open-loop tests.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Torque Limitation: The current system faces a torque limitation due to the HASEL actuators' constraints. This could be addressed by incorporating more actuators in parallel.
- Estimation Model Robustness: While the estimation model was effective for trained trajectories, further improvements are needed for handling new, unseen data. Advanced machine learning techniques are suggested for this.
- Controller Advancement: Introducing a model-based controller is proposed to enhance performance and enable control at higher frequencies.
- Third Yaw Axis Integration: The paper suggests exploring the integration of a third yaw axis to create a comprehensive three-DoF actuation mechanism, further enhancing the robotic shoulder's capabilities.
Summary updated