TrafficBots: Towards World Models for Autonomous Driving Simulation and Motion Prediction. ICRA 2023. You may also want to check out the updated version: https://github.com/zhejz/TrafficBotsV1.5
Hello, first of all I'd like to congratulate you on your work.
I'd like to try out your method, but I'm using Carla. Do you have a method for converting the Polylines Map you use into an OpenDrive Map, and vice versa?
I run the entire project with epoch 30(MinADE 1.612), It seems that the scene-center modeling method must be used to achieve faster convergence. I used the agent-center modeling method and found that it was difficult to converge. At the same time, the entire training process was slower than open-loop training and consumed more memory. I have some question,the paper only talks about the effect of using it on motion prediction problems, but the purpose of the world model is the planing tasks. Has the trafficbots been tested on nuplan planning task? Looking forward for your reply, thank you.
hey guys! I have read your paper, you claim that your work is closed-loop, have you considered introducing the nuPlan dataset, a benchmark that validates the planning performance of ego agent.
Hi,
in the bucket of waymo open dataset on google cloud : waymo_open_dataset_motion_v_1_2_0, I have a sub directory named : uncompressed, and in uncompressed I have 3 directories : lidar, scenario and tf_exemple. in the pack_h5.sh
file I see that for the training the data-dir is /cluster/scratch/zhejzhan/womd_scenario_v_1_2_0, does that mean you only use the directory scenario in the google cloud bucket (waymo_open_dataset_motion_v_1_2_0/uncompressed/scenario/) for the training, and we can ignore the lidar and tf_exemple directory?
Thank you for your attention