Empty action space

[yangyou95]

Thanks for providing this POMDP benchmark.

I would like to create a Roomba POMDP with a continuous action space, so I initialized Roomba POMDP as follows:

max_speed = 5.0 max_turn_rate = 1.0

pomdp = RoombaPOMDP(sensor=Bumper(), mdp=RoombaMDP(config=3, v_max=max_speed, om_max=max_turn_rate, sspace=sspace))

However, when I call action_space = actions(pomdp), it seems to return an empty object of RoombaPOMDPs.RoombaActions(), and I cannot use rand(action_space) to sample a random continuous action.

But according to the API documentation, actions(m::Union{MDP,POMDP}) should be able to return the entire action space of a (PO)MDP.

Please let me know if there is anything wrong during my initialization to create a continuous action space, thanks a lot!

[zsunberg]

It appears that rand is not implemented for the continuous action spaces. To see how to use a discrete action space, you can look here: https://github.com/sisl/RoombaPOMDPs.jl/blob/master/escape_roomba.ipynb

It looks like in general this package needs some maintenance:

1. upgrade to POMDPTools to help with compat issues
2. currently the tests are extremely slow - someone needs to track down why this is (possibly the Truncated distributions?)
3. add rand for continuous actions as @yangyou95 suggested

I don't have time to do this immediately, but I will ask my students.