Unity ML-Agent - Dynamically restrict agent actions | Official Unity Discord | Page 1

analog bay Mar 29, 2025, 3:47 PM

#

Hi hi, I'm working with Unity ML-Agents and using the Malbers Animal Controller to control a wolf character. I want to train an agent with a discrete action space, but I don't want all actions to always be available at every timestep.

My agent has mainly 2 actions: Speed mode (Walk, Sprint, Attack, Sneak), and Movement (forward/back/left/right)
If stamina is too low, the agent should only be allowed to walk and move.
If the agent is currently attacking, it should not be able to choose a movement at the same time.

I don't want the agent to learn from invalid actions. I want to prevent the policy from selecting those actions in the first place.

Any suggestions, or references would be helpful !
Thanks !!!

mild pollen Mar 31, 2025, 10:26 PM

#

This can be achieved using masking, have a look at this
https://unity-technologies.github.io/ml-agents/Learning-Environment-Design-Agents/#masking-discrete-actions

Agents - Unity ML-Agents Toolkit

The Unity Machine Learning Agents Toolkit (ML-Agents) is an open-source project that enables games and simulations to serve as environments for training intelligent agents.

analog bay Mar 31, 2025, 10:47 PM

#

Yes, that's interesting, thanks !
Is it more relevant to mask or to assign a negative reward for a forbidden action ?

#Unity ML-Agent - Dynamically restrict agent actions