Getting Started#

Quickstart colab notebook#

If you prefer to test the basic functionality of robomimic without installing anything locally, try the quickstart Colab notebook.

Running experiments#

We begin with a quick tutorial on downloading datasets and running experiments.

Before beginning, make sure you are at the base repo path:

$ cd {/path/to/robomimic}

Step 1: Download dataset#

Download the robosuite Lift (PH) dataset (see this link for more information on this dataset):

$ python robomimic/scripts/download_datasets.py --tasks lift --dataset_types ph

The dataset can be found at datasets/lift/ph/low_dim_v141.hdf5

Step 2: Launch experiment#

Now, we will run an experiment using train.py. In this case we would like to run behavior cloning (BC) for the lift dataset we just downloaded.

$ python robomimic/scripts/train.py --config robomimic/exps/templates/bc.json --dataset datasets/lift/ph/low_dim_v141.hdf5 --debug

Running quick sanity check experiments

Make sure to add the --debug flag to your experiments as a sanity check that your implementation works.

Resume functionality: If your training job fails due to any reason, you can re-launch your job with the additional --resume flag to resume training from the last saved epoch. This will resume training from the last.pth checkpoint in your output directory. Some points to note:

  1. While fine-tuning from a specified checkpoint (in config.experiment.ckpt_path) would load model weights from the checkpoint, resume functionality also loads the optimizer state.

  2. config.experiment.ckpt_path will be ignored if you are resuming a training job, i.e. last.pth will take precedence if --resume is passed.

Warning!

This example requires robosuite to be installed (under the v1.5.1 branch), but it can be run without robosuite by disabling rollouts in robomimic/exps/templates/bc.json: simply change the experiment.rollout.enabled flag to false.

Step 3: View experiment results#

After the script finishes, we can check the training outputs in the directory bc_trained_models/test. Experiment outputs comprise the following:

config.json               # config used for this experiment
logs/                     # experiment log files
  log.txt                    # terminal output
  tb/                        # tensorboard logs
videos/                   # videos of robot rollouts during training
models/                   # saved model checkpoints

The experiment results can be viewed using tensorboard:

$ tensorboard --logdir bc_trained_models/test --bind_all

Next steps#

Please refer to the remaining documentation sections. Some helpful suggestions on pages to view next: