Getting started with Dobb·E code

Getting familiar with running Dobb·E, including pretraining (or downloading) the Home Pretrained Representation model, and fine-tuning your own behavior policy.

In this section, we will help you get started with running all the software parts of Dobb·E. There are three major parts in the software component of Dobb-E, which are:

1. Processing data collected with the Stick,

2. Training a model on the collected data,

3. Deploying the model.

Optionally, you can pre-train your own model similar to how we trained the Home Pretrained Representations (HPR).

We will frequently refer to the "robot", which for this part would be the Intel NUC installed in the Hello Robot Stretch, and the "machine", which should be a beefier machine with GPU(s) where you can preprocess your data and train new models.

Also, ensure that you have mamba installed. Mamba is generally a faster but compatible alternative to conda, which we prefer.

Getting Started

• Clone the repository:

  git clone https://github.com/notmahi/dobb-e.git
  cd dobb-e/imitation-in-homes

• Set up the project environment:

  mamba env create -f conda_env.yaml

• Activate the environment:

  mamba activate home_robot

• Logging:

  • To enable logging, log in with a Weights and Biases (wandb) account:

    wandb login

  • Alternatively, disable logging altogether:

    export WANDB_MODE=disabled