search

mit-acl / cadrl_ros

ROS package for dynamic obstacle avoidance for ground robots trained with deep RL
573 stars 157 forks source link

The problem of running code #11

Closed BingHan0458 closed 3 years ago

BingHan0458 commented 3 years ago

hello, author, I try to run 'ga3c_cadrl_demo.py' and it prints 'action: [1.2 0.52359878]'. I understand that it means the next action of the host_agent and is that correct? And how to get the figures (Fig. 4 and Fig. 6) in you paper if I run the code 'network.py'?

BingHan0458 commented 3 years ago

And which models do network_01653000, network_01900000 and network_02360000 correspond to, respectively? How can I watch their structures? Thanks!

mfe7 commented 3 years ago

Yes, you're correct that is the action for the host_agent. If there are multiple agents in the environment, they would each compute their next action separately (presumably before any of the agents execute the action at that timestep).

To re-create Figs 4 and 6 from the IROS paper, it may be easier to use our gym environment repo. In particular, there's a small test suite script that I think runs those test cases. This might also be a useful repo if you're looking to compare algorithms across random scenarios.

Whichever network file is loaded by default is the GA3C-CADRL-10 that should perform the best. The other ones are either trained for fewer steps or might have had different settings, so it's probably best to ignore the other network files.