Add Mixture-of-Experts implementation

[Giulero]

This pull request introduces the implementation of Mixture of Experts (MoE) as actor.

---

New Features:

• Introduction of Mixture-of-Experts Models:
  • Added ActorMoE class, which implements a Mixture-of-Experts actor model. This model combines outputs from multiple expert networks using a gating network and softmax weights. (amp_rsl_rl/networks/ac_moe.py, amp_rsl_rl/networks/ac_moe.pyR1-R170)
  • Added ActorCriticMoE class, which extends the ActorMoE with a critic network for actor-critic reinforcement learning. It includes support for noise modeling and action distribution. (amp_rsl_rl/networks/ac_moe.py, amp_rsl_rl/networks/ac_moe.pyR1-R170)

Codebase Updates:

• Module Initialization Updates:
  • Updated __init__.py to include ActorMoE and ActorCriticMoE in the module's exports, ensuring they are accessible when importing the amp_rsl_rl/networks package. (amp_rsl_rl/networks/__init__.py, amp_rsl_rl/networks/init.pyR10-R13)

[Copilot]

Pull Request Overview

This pull request introduces a Mixture-of-Experts (MoE) implementation for actor networks in reinforcement learning. The implementation provides both standalone actor MoE and actor-critic MoE classes that combine outputs from multiple expert networks using a gating mechanism.

• Adds ActorMoE class implementing mixture-of-experts actor model with gating network
• Adds ActorCriticMoE class extending ActorMoE with critic network for actor-critic RL
• Updates module exports and ONNX exporter to support the new MoE architecture

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
amp_rsl_rl/networks/ac_moe.py	New file containing MoE implementation with ActorMoE and ActorCriticMoE classes
amp_rsl_rl/networks/__init__.py	Updates module exports to include new MoE classes
amp_rsl_rl/utils/exporter.py	Adds ONNX export support for ActorMoE architecture

Comments suppressed due to low confidence (2)

amp_rsl_rl/networks/ac_moe.py:9

• Class name 'MLP_net' violates Python naming conventions. It should be 'MLPNet' or 'MLP' following PascalCase convention for class names.

class MLP_net(nn.Sequential):

amp_rsl_rl/networks/ac_moe.py:108

• Using the incorrectly named 'MLP_net' class. This should be updated once the class name is fixed to follow proper naming conventions.

        self.critic = MLP_net(num_critic_obs, critic_hidden_dims, 1, act)

[Giulero]

Tested and working. Merging!