This repository contains the code for the paper "View-Invariant Policy Learning via Zero-Shot Novel View Synthesis". For more information, please see our project page.
First, create a directory to act as a workspace. I would also recommend creating a virtual environment at this point (e.g. via conda or mamba). This code is tested with Python 3.8.
Then, clone the ZeroNVS repo (https://github.com/s-tian/ZeroNVS) into your workspace, and follow the setup instructions in that repository's README, including pulling the zeronvs_diffusion submodule.
In the ZeroNVS directory, install the threestudio module:
cd threestudio
pip install -e .
cd ..
Clone this repo (https://github.com/s-tian/vista) into your workspace. Install the robomimic, robosuite, and vipl package as locally editable (note: robosuite and robomimic are also available on pip, but you need to install the local versions, as I've made edits to them):
cd robosuite
pip install -e .
cd ..
cd robomimic
pip install -e .
cd ..
cd vipl
pip install -e .
Next, install mimicgen_environments and robosuite_task_zoo, by navigating to your workspace directory and then doing:
git clone https://github.com/NVlabs/mimicgen_environments
cd mimicgen_environments
git checkout 45db4b35a5a79e82ca8a70ce1321f855498ca82c
pip install -e .
cd ..
git clone https://github.com/ARISE-Initiative/robosuite-task-zoo
cd robosuite_task_zoo
git checkout 74eab7f88214c21ca1ae8617c2b2f8d19718a9ed
pip install -e .
Lastly, downgrade mujoco as instructed in the mimicgen_environments README.
pip install mujoco==2.3.2
I've dumped the output of pip list and mamba list for working environment to this repo at pip_list.txt and mamba_list.txt for cross-reference.
Note that this environment is not used for the reprojection baseline due to conflicts with Pytorch3D. To create that environment, follow the same steps but do not install ZeroNVS, instead install pytorch3d==0.7.4.
Pretrained ZeroNVS models are available on HuggingFace here: https://huggingface.co/s-tian/VISTA_Data/tree/main/models.
To use them, download the models and config file, and then set the constants in vipl/vipl/utils/constants.py to point to the downloaded models and config on your local machine.
The general workflow of running simulated experiments is:
- Download source datasets for the task you want to run.
- Convert the source dataset to a dataset containing potentially augmented observations.
- Train a model on the potentially augmented dataset, and evaluate the model on held-out views of the target task.
The source datasets for this project are directly obtained from robomimic and mimicgen_environments. The source datasets are converted to augmented datasets using the models, e.g. zeronvs, offline. Each source dataset was originally obtained from the download scripts in robomimic or mimicgen_environments.
For convenience, we have uploaded the source datasets to HuggingFace here: https://huggingface.co/datasets/s-tian/VISTA_Data.
Conversion can be done by running the dataset_states_to_obs_zeronvs.py script in the provided robomimic package. For instance, let's look at the experiment in experiments/paper/exp_1/small_perturb_lift_zeronvs:
First, we run
python dataset_states_to_obs_zeronvs.py --dataset ../../../mimicgen_environments/datasets/vista/exp_1/small_perturb/lift/low_dim_v141.hdf5 --output_name random_cam_zeronvs.hdf5 --done_mode 2 --randomize_cam_range small_perturb --camera_names agentview robot0_eye_in_hand --camera_height 84 --camera_width 84 --compress --exclude-next-obs --randomize_cam --parse-iters 1 --camera_randomization_type zeronvs_lpips_guard
which uses the zeronvs_lpips_guard camera randomization method to augment the low_dim_v141.hdf5 dataset we downloaded in step 1 (you may need to change the dataset path to match your local setup).
Note that in experiments/paper/exp_1/small_perturb_lift_zeronvs, we break up this job into 10 commands to parallelize the process. When parallelizing, we provide a utility script merge_hdf5.py to merge the sliced hdf5 files after they are generated.
Continuing with the example in experiments/paper/exp_1/small_perturb_lift_zeronvs, we run
python robomimic/robomimic/scripts/train.py --config robomimic/robomimic/exps/vista/exp_1/small_perturb/lift/zeronvs.json
In the config file robomimic/robomimic/exps/vista/exp_1/small_perturb/lift/nvs.json, you'll see the dataset to train on is specified as
"data": "../mimicgen_environments/datasets/vista/exp_1/small_perturb/lift/random_cam_zeronvs.hdf5"
which you need to modify to point to the output of the previous step, which is by default generated in the same directory as your source dataset.
The training script uses robomimic conventions in general, but one detail to note is this entry:
"random_camera_params": {
"camera_name": "agentview",
"sampler_type": "small_perturb"
},
which specifies that a test environment will be created, with observations from the agentview camera randomized according to the small_perturb sampler type. This can be modified to test on other distributions of views (see vipl/vipl/utils/camera_pose_sampler.py for more details).