Multimodal end-to-end learning for continuous control of exoskeletons for versatile activities
Various examples are provided related to continuous control of exoskeletons. In one example, a method includes obtaining IMU sensor signals associated with an exoskeleton attached to a limb of a subject; generating an exoskeleton control signal in response to the IMU sensor signals, the exoskeleton control signal generated by a control policy neural network trained offline from the exoskeleton using musculoskeletal human modeling and exoskeletal modeling with dynamics randomization; and controlling joint torques of the exoskeleton exerted on the subject based upon the exoskeleton control signal.
1. A method, comprising:
obtaining inertial measurement unit (IMU) sensor signals associated with an exoskeleton attached to a limb of a subject;
generating an exoskeleton control signal in response to the IMU sensor signals, the exoskeleton control signal generated by a control policy neural network trained offline from the exoskeleton using musculoskeletal human modeling and exoskeletal modeling with dynamics randomization; and
controlling joint torques of the exoskeleton exerted on the subject based upon the exoskeleton control signal.
2. The method of claim 1 , wherein the IMU sensor signals comprise joint angle and joint angular velocity of the limb.
3. The method of claim 2 , wherein the limb is a leg of the subject and the joint torques are associated with a hip joint.
4. The method of claim 1 , wherein training of the control policy neural network comprises:
training a motion imitation neural network and a muscle coordination neural network to simulate muscle activation, the training of the motion imitation neural network and the muscle coordination neural network based upon simulated joint angle and joint angular velocity corresponding to the simulated muscle activation; and
simultaneously training the control policy neural network based upon the simulated joint angle and joint angular velocity.
5. The method of claim 4 , wherein the training of the motion imitation neural network, the muscle coordination neural network, and the control policy neural network comprise reinforcement training.
6. The method of claim 1 , wherein the exoskeleton comprises processing circuitry configured to control operation of the exoskeleton based upon the trained control policy neural network.
7. The method of claim 6 , wherein the processing circuitry is configured to low pass filter the exoskeleton control signal prior to controlling the joint torques via a PD controller.
8. The method of claim 1 , wherein the control policy neural network provides end-to-end continuous control of the exoskeleton over a range of operations.
9. The method of claim 8 , wherein the range of operations comprise walking, running and squatting.
10. An exoskeleton, comprising:
a support structure configured to interface with a limb of a user;
an inertial measurement unit (IMU) sensor configured to sense the limb of the user;
an actuator coupled to the support structure; and
processing circuitry configured to:
obtain IMU sensor signals associated with movement of the limb of the subject;
generate an exoskeleton control signal in response to the IMU sensor signals, the exoskeleton control signal generated by a control policy neural network trained offline from the exoskeleton using musculoskeletal human modeling and exoskeletal modeling with dynamics randomization; and
control joint torques of the exoskeleton exerted via the actuator based upon the exoskeleton control signal.
11. The exoskeleton of claim 10 , wherein the IMU sensor signals comprise joint angle and joint angular velocity of the limb.
12. The exoskeleton of claim 11 , wherein the limb is a leg of the user and the joint torques are associated with a hip joint.
13. The exoskeleton of claim 10 , wherein training of the control policy neural network comprises:
training a motion imitation neural network and a muscle coordination neural network to simulate muscle activation, the training of the motion imitation neural network and the muscle coordination neural network based upon simulated joint angle and joint angular velocity corresponding to the simulated muscle activation; and
simultaneously training the control policy neural network based upon the simulated joint angle and joint angular velocity.
14. The exoskeleton of claim 13 , wherein the training of the motion imitation neural network, the muscle coordination neural network, and the control policy neural network comprise reinforcement training.
15. The exoskeleton of claim 10 , wherein the processing circuitry is configured to control operation of the exoskeleton based upon the trained control policy neural network.
16. The exoskeleton of claim 15 , wherein the processing circuitry is configured to low pass filter the exoskeleton control signal prior to controlling the joint torques via a PD controller.
17. The exoskeleton of claim 10 , wherein the control policy neural network provides end-to-end continuous control of the exoskeleton over a range of operations.
18. The exoskeleton of claim 17 , wherein the range of operations comprise walking, running and squatting.