Neuromorphic Manipulator-Arm Controller For Object Insertion With Force Feedback
Object insertion in robotics is an active field of research since industry keeps seeking efficient ways to solve this problem. The programming of robot systems to handle object insertion tasks in modern manufacturing scenarios, e.g., a USB, D-SUB, or specialized industrial connectors or artefacts, is very challenging due to complex contact dynamics and friction. Using classical machine learning approaches with energy expensive GPUs and computing resources is not applicable for scenarios in mobile robotics. Therefore, an implementation on neuromorphic hardware and spiking neural networks will address this issue thanks to sparsity and high energy efficiency.
We use Intel’s neuromorphic chip codenamed Loihi to host an adaptive spiking controller on a standard industrial robotic arm equipped with haptic sensors. Our approach is to use spiking reinforcement learning to fine-tune the positioning and insertion movement with force feedback.
This project is happening in the frame of the Intel Neuromorphic Research Community.
The main contributions of this research will be a reinforcement learning algorithm for spiking neural networks and a hardware demonstrator with a simulated counterpart in the Neurorobotics Platform.
We believe that neuromorphic hardware can bring great improvements in terms of latency compared with inference on classical hardware, thanks to its sparse and massively parallel nature. This is in a robot control use case of critical importance since real time decision is a key factor. On the AI side, reinforcement learning in spiking neural networks is not a largely covered field of research and we expect interesting progress.
01.01.2021 - 31.12.2021