Does copo use "traffic light" and "communication" information as observation?

Question

Does copo use "traffic light" and "communication" information as observation?

Closed this issue 5 months ago · 5 comments

iamrice commented a year ago

iamrice commented a year ago

Thanks.

Answer 1 · 2023-03-20T07:26:07.000Z

I mean the algorithm version in the paper.

Answer 2 · 2023-03-20T08:28:01.000Z

No. They are completely useless in previous work.

Answer 3 · 2023-03-24T02:47:08.000Z

@pengzhenghao Hi, I found that the experiment result of the parking lot in the paper is different from that in 《MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning》

I notice that the paper of metadrive claims "the result is based on MetaDrive 0.2.5 and is different to previous paper [35] due to updates in MetaDrive.". But I am still curious what makes them differ so much?

Actually, I tried to implement ppo for metadrive, but i got a very poor success rate in parking lot. The success rate is about 0~0.2.

Thanks.

Answer 4 · 2024-01-29T21:01:33.000Z

i got a very poor success rate in parking lot. The success rate is about 0~0.2.

That's align with our benchmark result with 0.2.5: https://github.com/decisionforce/CoPO/tree/main/benchmarks/MetaDrive-0.2.5

I probably will run new benchmark experiments to see the performance. My suggestion is, you can stick with the MetaDrive version you have and run fair comparisions between different algorithms.