Powered Knee-Ankle Prosthesis

Legged Systems Group, Robotics Institute | Carnegie Mellon University | Ongoing

A number of control strategies have been developed for lower-limb prostheses, but current approaches still struggle with generalization and recognizing human intent. Reinforcement learning combined with learning in simulation is a promising way to address both but testing that idea requires iteration on control strategies quickly, and the CMU Leg’s control stack was built to run on a Simulink Real-Time target, not something suited to fast experimentation, especially for learning based approaches. So, I built the control stack on Python.

Hardware Overview

The motors on the leg are controlled by Elmo motor drives, one each for the knee and ankle. Ground reaction sensors and two encoders per joint feed a custom MCU. The drives and the MCU talk to a Jetson Orin over EtherCAT.

The EtherCAT interface is built on PySOEM, a Python library. Raw sensor data is fed into a state estimator, which outputs the state used by the high-level controller, the low-level controller then uses that output to command the motors.

Both the knee and ankle joints are series-elastic which means the joint and the motor are connected through a spring, which lets us measure torque for free, from the spring deflection rather than a dedicated torque sensor. The low-level controller here is velocity-based SEA control. It computes a velocity command and a feedforward torque from a dynamic model of the rotor, and sends the velocity command to the drives in velocity mode.

Control Testing

To stand, the leg first moves each joint to a zero position along a min-jerk trajectory, a smooth start-to-end motion profile that minimizes jerk (the rate of change of acceleration).

Early on, this control oscillated. Two things possibly fed into it 1) the control loop was only reliably hitting around 750 Hz instead of the target 1 kHz, which effectively inflated the velocity commands the control law was sending 2) there was a torque bias showing up even under no load, which was poisoning a few terms in the SEA control. Recalibrating the joint encoders against the motor-side encoders removed the bias, and after making the loop itself more efficient enough to hold a steady 1 kHz, the oscillations were gone.