Xiaoxiao Yang

e2e planning & RL

Focus on learning-based prediction & planning

• Proposed an end-to-end interaction-aware prediction and planning model with a hierarchical game among future trajectories of ego and agents. • In open-loop evaluation, the L2 error of planning trajectory decreased from 0.91m to 0.63m. The end-to-end prediction accuracy improved from 53.82% to 54.95%. • In the closed-loop evaluation of CARLA,  the close-loop metric (Driving Score) rose from 68.7 to 78.3 points.

• Designed and implemented a prediction model for cut-in scenes to solve the problem that the deployed main model does not predict the inconspicuous cut-in vehicles in time. • Proposed cutin-oriented features and transformer-based model for cut-in prediction task, conducted predictions in Frenet coordinate system, and improved the sensitivity and the robustness of prediction. • Deployed the prediction model on the Nvidia Orin platform, and its inference time is 4ms, with a precision of 0.93 and a recall of 0.84, effectively avoiding high-risk cut-in behaviors of other vehicles in advance.

• Optimization of VRU avoidance decision strategy ○ The original decision strategy of VRU avoidance is overly conservative and easily leads to emergency braking. Optimized the avoidance rules according to traffic rules and VRU interaction logic and nearly doubles the MPI. • Rule-based and model-based abnormal parking vehicle reasoning ○ Designed and implemented a rule-based reasoning module considering the static and dynamic environment, including agents history, maps, traffic lights, blind spots, etc. During the 700km road test, MPI for this issue increased from 70km to 150km.  ○ Designed and implemented a model-based reasoning module due to the increasing difficulty of rule maintenance. Collected abnormal parking vehicle datasets by a rule-based trigger, designed abnormal parking vehicle features and trained on the random forest algorithm, and obtained the precision of abnormal vehicle reasoning as 0.95 and the recall of 0.85.

Participated in the design and implementation of the evaluation system for AVP (Automated Valet Parking)  perception algorithm, including the introduction of new metrics, scenario classification and test set construction.

Vision4Robotics Lab, work with Prof. Changhong Fu

Outstanding Graduates of Tongji University