search

Thinklab-SJTU / Bench2Drive

[NeurIPS 2024 Datasets and Benchmarks Track] Closed-Loop E2E-AD Benchmark Enhanced by World Model RL Expert
Other
1.3k stars 85 forks source link

Inquiry on Route and Scenario Generation Process #26

Closed haibao-yu closed 4 months ago

haibao-yu commented 4 months ago

Hi, all:

Thank you for your outstanding work in comprehensively evaluating the end-to-end autonomous driving system using a closed-loop approach. I am curious about the methods you use to obtain the routes and traffic scenarios for generating the clips. Could you provide detailed information on the generation process?

Hope to receive your reply!

Best regards

jiaxiaosong1002 commented 4 months ago

@haibao-yu Thanks for your interests. Unfortunately, no. We do not plan to release the expert related code.

exiawsh commented 4 months ago

@jiaxiaosong1002 Hi, could you please share some details about the expert. Is the expert a rule-based model or RL agent (e.g. ,roach, think2drive)? Because the quality of the expert is important.

exiawsh commented 4 months ago

I also noticed that you mentioned PDM-lite. Are the results from the same b2d test set? Because the Driving score in PDM-lite is so high. image image

jiaxiaosong1002 commented 4 months ago

@exiawsh Hi, the expert model is Think2Drive, as stated in the paper.

As for the PDM-Lite, yes, it is on the Bench2Drive test set. The high score is because it is expert model with privileged inputs (GT boxes, map, etc). Other baselines are e2e student model with only raw sensor inputs.