Rewards tracking for multi-agent

[berttggg]

I am reading the code of rewards tracking for multi-agents.

And it seems that we are only interested in the first key's value, meaning the first agent only. Why? Or Anything I have missed?
Thank you.

skrl/skrl/multi_agents/torch/base.py

Line 305 in 7b090a1

_rewards = next(iter(rewards.values()))

[Toni-SM]

Hi @berttggg

Sorry for late response.

Yes, the current implementation uses only the reward from the first agent.

It was implemented using the Bi-DexHands environments for testing. Those environments define only one reward for all the task like Isaac Gym:
https://github.com/PKU-MARL/DexterousHands/blob/99c1e2a399fb084df5c02dbb5f6182d394fcd2e8/bidexhands/tasks/hand_base/base_task.py#L60-L61

In the next releases this implementation will be extended to a more generic form.

[berttggg]

I see, may I ask so far, the MAPPO or IPPO has any successful examples? How is the learning performance for the Bi-DexHands task?