Teaching Robots to be Fast Learners
A few months ago, I shared a vision for how energy-efficient, low latency neuromorphic computing can make assistive robotics a more practical and affordable tool to help people with upper body impairment perform tasks independently.
I’ve been working with our collaborators at the Neuro-Biomorphic Engineering Lab (NBEL) at Open University of Israel and Applied Brain Research to implement this vision. We are testing our hypothesis that the low latency of onboard adaptive control algorithms running on an energy-efficient neuromorphic processor can help robots move more precisely. These processors can reduce power consumption, reduce latency, and generate precise movement trajectories needed to make a wheel-chair mounted robot arm a practical and performant reality.
It’s one thing to make a robot arm move accurately in a simulation, where every robot joint and servo works exactly as expected. Similarly, accurate robot arm movement is easier in a perfectly stable, carefully arranged physical environment, controlled by a system that has complete knowledge of all the relevant objects and surfaces in that environment.
However, it’s a much harder challenge to make a practical robot work “in the wild,” where it’s subject to perturbations such as a changing center of mass when adding object weight, friction, and accounting for physical limitations at the joints, as well as needing to avoid obstacles. When the arm is carrying something like a bottle of water, where the object weight can shift around, accurate movement becomes even more difficult. For example, in the video above we can see that this causes the arm to unintentionally run into the table.
To solve these problems, two key algorithms required for robot control - inverse kinematics (IK) and Proportional Integral Derivative (PID) control - have to run faster while remaining energy-efficient. Given a target position in task space, inverse kinematics is used to compute an appropriate state in the robot’s configuration space. PID control applies responsive correction signals to a robot’s actuators, allowing it to reach its target accurately.
Conventional approaches to robot control, including IK and PID, traditionally require power-hungry hardware. We implemented these algorithms using spiking neural networks (SNNs), running on a neuromorphic processor, leveraging the energy efficiency and fast-learning capabilities of the neuromorphic hardware. We found that Neuromorphic implementations of IK and PID control outperform conventional control paradigms in terms of robustness to perturbations and adaptation to varying conditions.
In the future, we are planning to add computer vision capabilities to the robot arm using RGBD and and/or event-based cameras. These cameras provide enhanced visual information, such as depth and object location, which would enable better precision of robot arm movement. To learn more about our work in neurorobotics, check out our paper: Neuromorphic NEF-Based Inverse Kinematics and PID Control, visit our Accenture Labs page or Future Technologies R&D group page. For more information about this research contact me directly.