Much Ado About NoisingDispelling the Myths of Generative Robotic Control

Quick Example ​

# Install dependencies
uv sync --extras robomimic

# Train a flow matching policy on Robomimic Lift task
uv run examples/train_robomimic.py \
    task=lift_ph_state \
    network=chiunet \
    optimization.loss_type=flow \
    log.wandb_mode=online

Key Features ​

Multiple Training Objectives ​

This repository supports a variety of training objectives for generative behavior cloning:

• Flow Matching (flow): Standard continuous normalizing flow objective
• Regression (regression): Direct supervised learning baseline with the same architecture as flow matching.
• MIP (mip): Minimum Iterative Policy with two-step sampling (from Much Ado About Noising)
• TSD (tsd): Two-Stage Denoising (from Much Ado About Noising)
• CTM (ctm): Consistency Trajectory Model
• PSD (psd): Progressive Self-Distillation
• LSD (lsd): Lagrangian Self-Distillation (only support differentiable networks)

Flexible Network Architectures ​

Choose from multiple proven architectures:

• MLP: Enhanced MLP with Film conditioning, residual connection and layer normalization.
• ChiUNet: U-Net from Diffusion Policy (Chi et al.), good for smooth control tasks.
• ChiTransformer: Transformer architecture from Diffusion Policy, more expressive than ChiUNet but less stable.
• SudeepDiT: Diffusion Transformer (DiT) architecture, contains best practices from Sudeeep et al.
• RNN: LSTM/GRU-based recurrent networks, only for baseline purposes.

Comprehensive Task Support ​

Pre-configured environments and datasets:

• Robomimic: Manipulation tasks (Lift, Can, Square, Transport, Tool Hang)
• Kitchen: Multi-task kitchen environment
• PushT: 2D pushing task with multiple observation modalities

What's Next? ​

• Quick Start Guide - Get up and running in minutes
• Architecture & Design - Understand the framework internals
• Adding a New Task - Extend MIP to your own environments

Citation ​

If you use this repository in your research, please cite: