The policy does not learn in the unicycle mode

Shuijing725 / CrowdNav_Prediction_AttnGraph

[ICRA 2023] Intention Aware Robot Crowd Navigation with Attention-Based Interaction Graph

https://sites.google.com/view/intention-aware-crowdnav/home

MIT License

147 stars 28 forks source link

The policy does not learn in the unicycle mode #18

Open ehsankf opened 2 weeks ago

ehsankf commented 2 weeks ago

Hi,

I trained the model with the unicycle kinematic and using 20 humans as obstacles and I noticed that the policy does not learn. Is there any requirement for the number of humans with unicycle kinematic? I believe the interval (-0.06, 0.06) for the change of angle i.e., \Delta \theta is too small for the agent to navigate this dense environment.

Thanks.

Shuijing725 commented 2 weeks ago

If you have a strong computer, you can increase the number of parallel environments by changing args.num_processes and the number of training steps args.num_env_steps in arguments.py . Otherwise, you can reduce the task difficulty by fixing the starting and goal positions of robot and/or humans.