Find in Library

High-speed path planning is a real-time task that requires vehicles to operate at the edge of their performance limits and holds significant practical importance. However, the significant differences between the simulated training environments and real-world environments pose a challenge, leading to performance discrepancies between the simulator and real-world applications, thus impacting effective high-speed path planning. To address this issue, we propose a novel reinforcement learning network that identifies the environment by extracting scene graphs. We utilize a forward dynamics environment simulation network to mitigate the discrepancies between the simulator and reality and enhance the reward function to expedite training convergence. Ultimately, our algorithm demonstrates excellent performance in both physical simulations and real-world scenarios, achieving high-speed path planning. Compared to previous reinforcement learning path planning algorithms, we have increased the race completion success rate by more than 20.83%, proving the effectiveness of our algorithm.