A small question about parallel RL traininng for social navigation

[JinJinSang]

Dear Shuijing,

Thx so much for your efforts on collecting and then sharing the information of these papers in the field of social navigation. It has saved me a lot of times. I have seen that people mainly focus on either the design choice of the network or the simulator. However, most of the RL based methods in this field are still using non-parallel RL (only a few papers use parallel RL but not in large scale because they mainly do parallelization on CPU).

Do you think it's necessary for this field to incorporate large-scale parallel training into the training or do you know any paper in this field achieves large-scale RL training?

Best,

[Shuijing725]

I agree about your observation on lack of GPU social navigation simulators and the need for GPU simulators. In lots of papers (including mine), people and static obstacles are abstracted into geometric shapes (circles, polygons, etc), so the amount of compute can be handled by parallel CPU threads. However, some cues that matter for social navigation are lost, including but not limited to facial expressions and whole-body motions of pedestrians, curbs, lanes, and signages. It would be great if we have something like an IsaacSim social navigation simulator.

[JinJinSang]

Thx for your quick reply Shuijing !!

May I ask one more question, do you know any customized version CrowdSim that supports irregular obstacles as well as the “human occlusion” check among all the human and static obstacles ?