Research On Decision And Planning Of Autonomous Vehicle In Dynamic Interaction

The decision-making and planning system for self-driving cars is crucial for enhancing vehicle safety and efficiency in the future.The decision-making and planning process inevitably requires dynamic interaction with other traffic participants.However,three key technical problems still exist with the dynamic interaction-based decision-making and planning approach.First,due to the complexity and variability of dynamic scenes,the trajectory of surrounding vehicles has a high degree of uncertainty,and traditional single-modal trajectory prediction methods cannot fully represent it.Multi-modal trajectory prediction methods also require a concise and unified multi-modal integrated prediction framework.Second,traditional decisionmaking and planning methods do not consider the bidirectional interaction characteristics between the host vehicle and surrounding vehicles,resulting in overly conservative methods in dynamic scene interactions that perform well in static scenes.Third,decision-making and planning systems often require dynamic interactions in diverse scenes,which demand high levels of generalization.To address the above three issues,this article proposes a dynamic interaction decision-making and planning method for self-driving cars in multiple scenarios based on structured road environments and surrounding traffic vehicles.To address the problem of insufficient representation of vehicle trajectory uncertainty by traditional single-modal trajectory prediction models,this article proposes a multi-modal vehicle trajectory prediction method based on spatiotemporal generation representation.This method integrates multiple modes of probability generation models for unified representation with diversity and avoids the problem of missing trajectory caused by statistical averaging of trajectory uncertainty.The residual network and the long short-term memory network are used to compensate for the insufficient time-space representation capability of traditional generation models.The confidence of multi-modal trajectory prediction is further optimized using the modal prediction model,reducing trajectory prediction errors caused by confidence bias.To address the problem of conservative decision-making caused by traditional one-way interaction decision-making and planning methods’ inability to capture the impact of host vehicle decisions on surrounding traffic vehicles,this article proposes a dynamic interaction decision-making and planning method based on multi-agent reinforcement learning explicit games.This method uses the theory of Nash equilibrium to solve the paradoxical problem of optimal bidirectional interaction strategy and converts it into the convergence problem of multiagent reinforcement learning itself using the convergence condition of Nash equilibrium,relying on the recursive explicit game form of multi-agent to solve the dynamic Nash equilibrium.To address the scalability problem inherent in the multi-agent explicit game model,this article proposes a bidirectional interaction target discrimination method based on interaction value analysis.This method can prioritize the high-value interaction vehicles based on value preferences and indirectly improve the algorithm’s scalability.To address the generalization problem of multi-agent models in multiple scenarios.We construct a meta reinforcement learning policy model and integrate multiple policy model parameters using the meta-model to solve the catastrophic forgetting problem of generalization training.We utilize pre-training and sampling meta-training based on Dirichlet multi-scenario distribution to train the model parameters in different scenarios,enabling online applications while effectively suppressing generalization training problems caused by scenario differences.We conducted experiments on simulation platforms and real-world scenario datasets for each individual method and the overall method.The experimental results demonstrate the safety and efficiency of our approach.This paper provides a systematic approach and methodology for the research of dynamic interactive autonomous driving decision-making and planning systems,and provides theoretical guidance and technical support for the practical application of autonomous driving vehicles.