Home

Imitation Learning for Any Robot Policy.

---

VersatIL is a composable PyTorch framework for training robot policies through Imitation Learning. It decouples Data, Algorithm, Architecture, and Loss so you can build, benchmark, and deploy any policy from config alone.

The key features are:

• Composable: A policy is assembled from an encoding pipeline, an algorithm, an action decoder, and a loss module. Swap any component with a config change.
• Any encoder: Use any vision backbone from timm, any language model from HF Transformers, or custom geometric encoders for depth -- fuse them with attention, MLP, or concatenation.
• Any algorithm: BehavioralCloning, Diffusion, FlowMatching, and VariationalAlgorithm with pluggable prior-posterior schemes.
• Action decoders: ACT, DiT, GPT, LACT, MoDE-ACT, PhaseACT, AutoregressiveVLADecoder, OpenVLAOFTDecoder, Pi0Decoder, and SmolVLADecoder.
• Any dataset: Ingest CSV, HDF5, or LeRobot formats into Zarr. Define observation/action spaces per task. Normalize, augment, and tokenize automatically.
• Config-driven: Hydra + typed OmegaConf dataclasses. Every experiment is fully reproducible from a single YAML file. Errors are caught at startup, not mid-training.
• Inference-ready: Pluggable transport protocols (ZMQ), observation buffering, temporal aggregation, structured actions. One client for simulation and hardware. Post-training compression (pruning + quantization via torchao) for deployment on resource-constrained hardware.
• Tested: >90% test coverage. Ruff formatting and linting enforced via CI/CD and pre-commit hooks.

---

Getting Started

• Installation

  ---

  Set up Python, CUDA, and VersatIL with conda + uv.

  Install

• First Training Run

  ---

  Train a policy, override configs from CLI, resume from checkpoints.

  Train

• Configuration

  ---

  Hydra composition, OmegaConf typed configs, interpolation.

  Configure

Architecture

• Overview

  ---

  Policy = EncodingPipeline + Algorithm + ActionDecoder + Loss.

  Overview

• Encoding Pipeline

  ---

  Multi-modal encoders (RGB, depth, language, VLM) and fusion modules.

  Encoders

• Algorithms

  ---

  BC, Diffusion, FlowMatching, and VariationalAlgorithm.

  Algorithms

• Action Decoders

  ---

  Decoder architectures: ACT, DiT, GPT, AutoregressiveVLADecoder, OpenVLAOFTDecoder, Pi0Decoder, and SmolVLADecoder.

  Decoders

• Data Pipeline

  ---

  Zarr ingestion, normalization, augmentation, tokenization.

  Data

• Inference

  ---

  Transport protocols, preprocessing, temporal aggregation.

  Inference

• Explainability

  ---

  Grad-CAM, Grad-CAM++, and Ablation-CAM heatmaps for any trained policy.

  Explainability

• Quantization

  ---

  Quantization-aware training and eager/PT2E workflows via torchao.

  Quantization

• Post-Training Compression

  ---

  Pruning, quantization, and export for compressed inference.

  Compression

Reference

• API Reference

  ---

  Auto-generated from source code docstrings.

  API

• Changelog

  ---

  Release history and notable changes.

  Changelog