Behavior Retrieval

This repository contains the code for BehaviorRetrieval, a few-shot imitation learning method that queries unlabeled datasets.

Installing

1. Install python 3.7
2. Install torch + torchvision
3. Install robosuite (pip install robosuite or install from source). You also need to install mujoco_py.
4. Install dependencies in requirements.txt (covers robomimic and roboverse dependencies)
5. Install robomimic by using pip install -e . inside the robomimic folder
6. (If you want to run Office) Install roboverse by using pip install -e. inside the roboverse folder

Running

Configurations

Find all the configurations for training in configs/. We follow the Robomimic convention of keeping hyperparameters in the .json files. We have special office configuraitons for the Office task due to differences of the Roboverse environment.

Collecting Data

• Can task: use the paired data provided by Robomimic: download
• Square task: use the MachinePolicy to collect demonstrations. Read the script in run_trained_agent.sh for more information
• Office task: use the scripted_collect.sh in the roboverse/scripts folder. Use utils/roboverse_to_robomimic.py to convert the demo format to the one used by our codebase

Training The Embedder

Use train_embedder.py, which can handle contrastive and VAE embeders. For configs, see train_embedder.sh

Running BehaviorRetrieval

Use run_weighted_BC.py, which runs BehaviorRetieval and our baselines. For example configs, see run_weighted_BC.sh

To train with vanilla BC, use train.py. We use this to pretrain the model for the HG-DAGGER experiments. See example config in train.sh.

To run HG-DAGGER with BehaviorRetrieval, use run_weighted_corrections.py. See run_weighted_corrections.sh for an example config.

Evaluating BehaviorRetrieval and Baselines

Use run_trained_agent.py to run evals. See run_trained_agent.sh for configs.

Real Robot Experiments

If the robot interfaces with the Gym environment, this codebase works with real robots out of the box.

Visualizing Results

Use visualizing_embeddings.py to compute a T-SNE or PCA visualization of the embeddings. Use embedding_analysis.py to compute a plot of how similarity changes through an episode (like Figure 8 in our paper)