Learning Inverse Dynamics Using Brain-Inspired Computational Principles on Neuromorphic Hardware

Background and Research Motivation

In the modern field of robotics, there is great potential for low-latency neuromorphic processing systems enabling autonomous artificial agents. However, the variability and low precision of current hardware foundations present significant challenges for achieving stable and reliable performance. To address these challenges, researchers have adopted brain-inspired computational principles, such as triplet spike-timing dependent plasticity, disinhibition mechanisms based on basal ganglia, and cooperative-competitive networks, applying these technologies to motion control.

This study demonstrates the feasibility of this approach by presenting a hardware spiking neural network (SNN) implemented using a mixed-signal neuromorphic processor for online learning of inverse kinematics for a two-joint robotic arm. Ultimately, the system can use noisy silicon neurons to achieve low-latency control with an accuracy of 97.93%, network latency of 33.96 milliseconds, system latency of 102.1 milliseconds, and a power consumption estimate of 26.92 microwatts during the inference phase.

Paper Source

This paper was written by Jingyue Zhao, Marco Monforte, Giacomo Indiveri, Chiara Bartolozzi, and Elisa Donati, who are from the University of Zurich, the Swiss Federal Institute of Technology Zurich (ETH Zurich), and the Italian Institute of Technology. The paper was published in 2023 in the journal “npj Robotics.”

Detailed Research Process

a) Research Workflow

The research process includes the following parts:

1. Data Collection and System Training: Training data was generated by controlling the shoulder pitch and elbow joints of the iCub robot in a simulated environment.
2. Model Design: A series of neural populations were designed to encode the target Cartesian coordinates of the end-effector and the corresponding joint angles.
3. Weight Training: On-chip SNN weights were trained using a “computer in the loop” approach to account for system imperfections, implementing learning rules based on spike timing.
4. Inverse Kinematics Solving: Synapses between two hidden layers were trained to learn the correct mapping strategy, introducing disinhibition mechanisms inspired by the basal ganglia and recurrent connections.
5. Robustness Verification and Low Power Testing: The robustness and reliability of the system were tested using noisy neurons.

b) Main Research Results

1. Inverse Kinematics Learning: By training the shoulder pitch and elbow joints, the SNN can drive the two joints in real-time to allow the end-effector to continuously reach target points in 2D space.
2. Low Latency and Low Power Consumption: In continuous target-reaching tasks, the SNN exhibited a network latency of 33.96 milliseconds and a chip power consumption of 26.92 microwatts.
3. Role of Disinhibition in Training: The study demonstrated how the disinhibition mechanism helps noisy silicon neurons form stable spiking patterns during the training phase to learn inverse kinematics.

c) Research Conclusions and Significance

This study shows that specific computational principles (such as disinhibition mechanisms and triplet spike-timing dependent plasticity) have important applications in solving complex engineering tasks and can draw on the design of biological neural systems in neuromorphic computing. The research provides strong evidence for the design of end-to-end spiking robotic control systems, marking an important step towards the development of efficient, low-power autonomous robotic platforms.

d) Research Highlights

1. High Accuracy and Low Latency: The inverse kinematics solver developed in this study achieved a low network latency of 33.96 milliseconds with an accuracy of 97.93%.
2. Biologically Inspired Innovative Mechanisms: The study used a basal ganglia inspired disinhibition mechanism to solve the problem of selecting among multiple possible solutions, demonstrating its effectiveness in coordinating motion tasks.
3. Low Power Solution: A method using low-power SNN for robotic motion control was proposed, showing its potential in practical engineering applications.

e) Additional Information

Future research directions include:

1. Extension to Multiple Degrees of Freedom: More neurons can be used in the future to increase the task space, extending the end-effector space from 2D to 3D, and expanding the joint configuration space.
2. Higher Neural Coding Resolution: Increasing the encoding resolution to reduce discretization errors.
3. Comprehensive Neural Architecture Construction: Using event-driven sensors and single-joint low-level controllers to establish a full spiking control process from high-level controllers to low-level execution.

This research provides a solid foundation for the development of future autonomous robotic systems, with potential applications including adaptive robotic control, low-power wearable devices, and efficient biomedical equipment.

Conclusion

This paper extends the application of neuromorphic computing in the field of robotic motion control, demonstrating an effective solution for the inverse kinematics problem through inspiration from biological neural systems. The study achieves a balance of low latency, low power consumption, and high accuracy in complex engineering tasks, providing an important reference for the development of future autonomous robotic platforms.