Installation

Requirements

Supported Operating Systems

  • Ubuntu 20.04 or newer is required.

  • Other Linux distributions and macOS may work but are not officially supported.

  • Windows is not supported.

GPU Requirements

  • An NVIDIA GPU is required. Minimum recommended GPU/VRAM for each use case:

    Mode

    Memory (VRAM)

    Example GPU

    Inference

    > 8 GB

    RTX 3090

    Training

    > 70 GB

    A100 (80GB) / H100

  • For most purposes, training and inference require NVIDIA GPUs with recent CUDA support (CUDA 11+, commonly available with driver version 450+).

  • Multi-GPU setups (A100, H100, etc.) should be used for large-scale training.

Installation with PyPI

You can install OpenTau directly from PyPI using pip:

pip install opentau

To install with extra dependencies (e.g., dev, libero, robocasa, urdf, trt), use brackets:

pip install opentau[dev,libero,robocasa,urdf,trt]

Installation with Source Code

Download Source Code

Download the source code:

git clone https://github.com/TensorAuto/OpenTau.git
cd OpenTau

Environment Setup

We recommend using uv (>= 0.8.4) for fast and simple Python dependency management.

  1. Install uv Follow the official uv installation instructions. OpenTau requires uv >= 0.8.4 because pyproject.toml uses [tool.uv].extra-build-dependencies (introduced in 0.8.4) to inject cmake into the PEP 517 build isolation of egl-probe (a libero extra dependency), so its sdist builds without a system cmake and on CMake 4. Older uv is rejected via required-version.

  2. Install dependencies Sync all required dependencies. To install all extras (they all co-resolve into a single environment):

    uv sync --all-extras

    To install specific extras (e.g., dev, libero):

    uv sync --extra dev --extra libero

  3. Activate the virtual environment

    source .venv/bin/activate

Docker Installation (Optional)

You can also use Docker to install and run OpenTau.

  1. Build the Docker image

    docker build -t opentau .

  2. Run the Docker container

    docker run -it --gpus all opentau /bin/bash

    Note: The --gpus all flag requires the NVIDIA Container Toolkit.

Experiment Tracking

To use Weights and Biases for experiment tracking, log in with:

wandb login

Distributed Training Configuration

Configure accelerate for your distributed training setup:

accelerate config

This will create an accelerate config file at ~/.cache/huggingface/accelerate/default_config.yaml. The recommended setup for models that fit in GPU memory (including the pi05 reference policy) is plain DDP with bf16 mixed precision. For an example, see configs/examples/accelerate_ddp_config.yaml.

A DeepSpeed ZeRO-2 config is also available at configs/examples/accelerate_deepspeed_config.yaml for memory-constrained scenarios (very large models, long sequences), but note that it can be significantly slower than DDP on mid-sized policies with many small parameter tensors due to per-parameter gradient-reduce hooks. See issue #177 for benchmarks.