Mapless Motion Planning for Autonomous Racing using Sim-to-Real Deep Reinforcement Learning скачать в хорошем качестве

Mapless Motion Planning for Autonomous Racing using Sim-to-Real Deep Reinforcement Learning

Autonomous racing requires quick processing at high speeds that has potential for technology transfer to other fields. Contemporary approaches utilize detailed maps which inhibits generalization to new, previously unseen racetracks. Motion planning at high speeds without prior maps is a challenging problem due to the highly nonlinear physics at a vehicle's tire and acceleration limits, that require significant computation resources with conventional optimization and sampling methods, which can be infeasible for real-time execution in embedded computers. This video presents a DRL method for mapless racing motion planning, trained in simulation over 20,000,000 steps in 48 wall-clock hours, transferred zero-shot to multiple new real-world racetrack configurations. A sparse reward and a training environment with a transition probability to facilitate Out-of-Distribution (OOD) parameterization are utilized, to train a compact, computationally efficient end-to-end Artificial Neural Network (ANN). Zero-shot sim-to-real transfer is achieved with a combination of high-fidelity simulation assets and a reward component to mitigate differences between simulation and real-world physics.