protomotions.agents.masked_mimic package

MaskedMimic implementation for versatile motion control.

This package implements the MaskedMimic algorithm which learns to reconstruct expert tracker actions from partial observations. Trains on data from a full-body motion tracker while randomly masking observations.

Key Components:

• MaskedMimic: Main MaskedMimic agent

• MaskedMimicModel: Model with optional VAE

• MaskedMimicAgentConfig: Configuration

Training Process:

1. Phase 1: Train expert full-body tracker (separate)

2. Phase 2: Train MaskedMimic to imitate expert with masked observations

Example

>>> from protomotions.agents.masked_mimic.agent import MaskedMimic
>>> config.expert_model_path = "results/expert_tracker/"
>>> agent = MaskedMimic(fabric, env, config)
>>> agent.train()

Reference:

Tessler et al. “MaskedMimic: Unified Physics-Based Character Control Through Masked Motion Inpainting” (2024)

Submodules

• protomotions.agents.masked_mimic.agent module
  • MaskedMimic
    • expert_model
    • vae_noise
    • env
    • model
    • __init__()
    • config
    • setup()
    • create_model()
    • create_optimizers()
    • reset_vae_noise()
    • post_env_step_modifications()
    • add_agent_info_to_obs()
    • load_parameters()
    • register_algorithm_experience_buffer_keys()
    • collect_rollout_step()
    • perform_optimization_step()
    • masked_mimic_step()
    • calculate_extra_loss()
    • get_state_dict()
• protomotions.agents.masked_mimic.config module
  • KLDScheduleConfig
    • init_kld_coeff
    • end_kld_coeff
    • start_epoch
    • end_epoch
    • __init__()
  • VaeNoiseType
    • NORMAL
    • UNIFORM
    • ZEROS
    • from_str()
  • VaeConfig
    • kld_schedule
    • vae_latent_dim
    • vae_noise_type
    • __init__()
  • FeedForwardModelConfig
    • trunk
    • __init__()
  • MaskedMimicModelConfig
    • encoder
    • prior
    • trunk
    • vae
    • optimizer
    • __init__()
  • MaskedMimicAgentConfig
    • model
    • expert_model_path
    • __init__()
• protomotions.agents.masked_mimic.model module
  • FeedForwardModel
    • __init__()
    • forward()
  • MaskedMimicModel
    • config
    • __init__()
    • reparameterization()
    • forward()
    • kl_loss()