Neuromorphic computing has potential in future autonomous systems as it provides unique advantages over traditional computing methods. Despite this potential, there have been few real-world demonstrations of neuromorphic computing in this space. Led by Robert Patton, Group Lead of the Learning Systems Group, a team at ORNL has designed a workflow for developing a spiking neural network for real-world autonomous racing. The team used F1Tenth, a one tenth scale racing competition for autonomous agents, as a platform for demonstrating this workflow. Evolutionary Optimization for Neuromorphic Systems (EONS) was used in conjunction with an OpenAI Gym environment provided by F1Tenth to train and test the model. The model uses a LIDAR sensor as input and it outputs steering and speed controls to drive the car. Initial results have shown that trained models can perform consistently well on three of the five training tracks, while fewer models perform well on the most difficult two, all while keeping a relatively small footprint due to having sparse networks. Future work includes further refinement of the model and the training process, as well as deployment to the physical car via µCaspian on an FPGA board. [Catherine D. Schuman et al., “Neuromorphic Computing for Autonomous Racing,” International Conference on Neuromorphic Systems, 2021. DOI: TBD]
