[Paper Review] Learning to Control Highly Accelerated Ballistic Movements on Muscular Robots

Dieter Büchler, Roberto Calandra+1 more|arXiv (Cornell University)|Apr 7, 2019

Muscle activation and electromyography studies4 citations

TL;DR

This paper presents a 4-DOF lightweight robotic arm actuated by antagonistic pneumatic artificial muscles (PAMs) that achieves extreme joint accelerations up to 28,000°s⁻² while safely exploring high-speed control via Bayesian optimization. The system enables direct, safe learning of fast ballistic trajectories through inherent PAM compliance, co-contraction control, and mechanical design minimizing friction and cable deflection, outperforming prior PAM-driven robots in speed and tracking accuracy.

ABSTRACT

High-speed and high-acceleration movements are inherently hard to control. Applying learning to the control of such motions on anthropomorphic robot arms can improve the accuracy of the control but might damage the system. The inherent exploration of learning approaches can lead to instabilities and the robot reaching joint limits at high speeds. Having hardware that enables safe exploration of high-speed and high-acceleration movements is therefore desirable. To address this issue, we propose to use robots actuated by Pneumatic Artificial Muscles (PAMs). In this paper, we present a four degrees of freedom (DoFs) robot arm that reaches high joint angle accelerations of up to 28000 deg/s^2 while avoiding dangerous joint limits thanks to the antagonistic actuation and limits on the air pressure ranges. With this robot arm, we are able to tune control parameters using Bayesian optimization directly on the hardware without additional safety considerations. The achieved tracking performance on a fast trajectory exceeds previous results on comparable PAM-driven robots. We also show that our system can be controlled well on slow trajectories with PID controllers due to careful construction considerations such as minimal bending of cables, lightweight kinematics and minimal contact between PAMs and PAMs with the links. Finally, we propose a novel technique to control the the co-contraction of antagonistic muscle pairs. Experimental results illustrate that choosing the optimal co-contraction level is vital to reach better tracking performance. Through the use of PAM-driven robots and learning, we do a small step towards the future development of robots capable of more human-like motions.

Motivation & Objective

• To enable safe, direct learning of high-acceleration ballistic movements on physical robotic hardware.
• To overcome the instability and hardware damage risks associated with exploration in high-speed control using traditional motor-driven arms.
• To design a PAM-actuated robot with minimal mechanical friction and optimized cable routing to improve controllability.
• To demonstrate that co-contraction control of antagonistic PAM pairs significantly enhances trajectory tracking performance.
• To validate that Bayesian optimization can be applied directly on hardware without additional safety constraints due to inherent system compliance.

Proposed method

• Design of a 4-DOF robotic arm using eight antagonistic PAMs, with each joint actuated by two PAMs in opposition.
• Implementation of a hardware system with low-friction cable routing using Bowden cables and PAMs mounted to minimize deflection and bending.
• Use of proportional valves and an FPGA-based control system to regulate PAM pressure at 100 Hz with 100 kHz encoder feedback.
• Application of Bayesian optimization (BO) to tune PID controllers with feedforward terms directly on hardware, using predefined pressure and parameter limits as safety constraints.
• Introduction of a novel co-contraction control technique to optimize muscle pair activation levels for improved tracking performance.
• Estimation of the Pareto front in the objective space (position and velocity error) to analyze the impact of co-contraction levels on control performance.

Experimental results

Research questions

• RQ1 Can PAM-driven robotic arms safely explore high-acceleration movements using learning-based control without external safety mechanisms?
• RQ2 How does co-contraction level affect trajectory tracking performance in a nonlinear, high-acceleration robotic system?
• RQ3 To what extent can a simple PID controller with feedforward terms achieve high-precision tracking on a PAM-actuated arm with minimal mechanical complexity?
• RQ4 Does mechanical design minimizing cable deflection and PAM-to-structure contact improve controllability of fast movements?
• RQ5 Can Bayesian optimization be effectively applied directly on hardware for high-speed trajectory control in a nonlinear system?

Key findings

• The robot achieved a peak joint space acceleration of 28,000°s⁻² and a task space acceleration of 200 m/s², exceeding the Barrett Wam arm by 10x in acceleration and 4x in velocity.
• The system sustained high forces from rapid accelerations without mechanical damage, enabling safe exploration during learning.
• Bayesian optimization successfully tuned PID controllers with feedforward terms directly on hardware, achieving superior tracking performance on fast trajectories.
• Experiments showed that trials close to the estimated Pareto front shared similar co-contraction levels, indicating that optimal co-contraction is critical for performance.
• The robot achieved precise tracking on slow trajectories using only manually tuned PID controllers, demonstrating the effectiveness of mechanical design in reducing nonlinearity.
• The absence of diverse co-contraction levels near the Pareto front suggests that linear control can effectively manage the nonlinear dynamics when co-contraction is properly selected.