- the script
WolfAgent.cs(in Assets) needs to be debugged so that it doesn't spit out errors during training. - the script should be so that whenever the wolf noclips from the platform, it gets moved back to the place of origin. That part doesn't currently work.
Solution:
root.localPositionis position relative to the parent - in our case to WolfAgent, it should be relative to Training Area instead - when starting up the training process, I (Maciek) get an error after a ~minute: ![[Pasted image 20230417181924.png]] If others get this error as well, it needs to be resolved.
- Think about a more lightweight ragdoll, so that the model doesn't have such a hard time working with physics
- need to figure out how to enable Soft Actor-Critic. Supposedly ML-Agents has an
- need to make double extra extra sure that the learning process is working as it's supposed to:
- what should be the settings for
BehaviorParameterscomponent ofWolfAgent? - are there any components we should implement? The docs mention
DecisionRequester,StatsRecorderetc. Which one of these do we need? Which ones we don't need?
- what should be the settings for
- On a more theoretical level, we need to think how to motivate it to stand on its own two legs. I think this is basically 95% of the challenge.