SpawnNet: Learning Generalizable Visuomotor Skills from Pre-trained Networks

Xingyu Lin*, John So*, Sashwat Mahalingam, Fangchen Liu, Pieter Abbeel

[paper]() | website

Setup

For reproducibility, we provide steps and checkpoints to reproduce our simulation DAgger experiments. For real-world BC experiments, we additionally provide a dataset for visualization and BC training. For more information regarding the real-world setup, see Real World Experiments.

Create the environment:

cd spawnnet
conda create -n spawnnet python==3.8.13 pip
conda activate spawnnet
sudo apt install ffmpeg
pip install -r requirements.txt

# baselines
pip install git+https://github.com/ir413/mvp # MVP
pip install git+https://github.com/facebookresearch/r3m # R3M

Modify prepare.sh

prepare.sh is a file used to set up the necessary environment variables and library paths. You must modify prepare.sh as described in the file's comments. Make sure to source prepare.sh once completed:

. prepare.sh

Simulation Setup

We run our simulation experiments using IsaacGym with tasks lifted from RLAfford. To set up the IsaacGym environments developed by RLAfford, follow the instructions given at RLAfford/README.md.

PIP Troubleshooting

If when installing any packages, you get a PIP error that extra_requires must be a dictionary, consider changing your setuptools version through pip install setuptools==65.5.0, then rerunning.

NVIDIA Troubleshooting: Driver Mismatches/Issues

If you encounter issues with a driver mismatch between your CUDA and NVIDIA Drivers, consider these two steps:

1. Consider adding the LD_LIBRARY_PATH, PATH, and CUDA_HOME changes from prepare.sh to your bash profile ~/.bashrc. This may need a terminal and/or system restart.
   • Typically, legacy CUDA versions may interfere with graphics processes before the variables are updated in prepare.sh. This step is meant to resolve that issue.
2. Ensure that your CUDA version >= 11.7 (nvcc --version), and that you have a compatible NVIDIA-Driver version. Any re-installations may require a system reboot.

Repository Structure

conf

We use hydra to manage experiments. Configs correspond exactly to the module in simple_bc.

gdict

This is a library for storing dictionaries of tensors; supports array-like indexing and slicing, and dictionary-like key indexing. Extracted from ManiSkill2-Learn

simple_bc

This can be roughly split into 3 modules:

1. dataset: this loads preprocessed hdf5 files into GDict structs.
2. encoder: this processes inputs into latent vectors.
3. policy: these are learning algorithms to output actions.

The network modules all follow interfaces defined in _interfaces. To add a new network, implement the abstract methods in each interface (see encoder/impala.py for an example), add the network to the module __init__.py file (see encoder/__init__.py) and define a hydra configuration in root's conf (see conf/encoder/impala.yaml).

Additionally, we provide scripts for training and evaluating policies under train.py and eval.py.

Simulation Experiments

Note: Before running experiments in a terminal, be sure to source prepare.sh first.

Training

There are two tasks in simulation, Open Drawer and Open Door. The IsaacGym configurations for both tasks can be found under RLAfford/cfg/open_door_expert.yaml and RLAfford/cfg/open_drawer_expert.yaml, respectively.

After setting up everything, set only WHICH_GPUS if in non-SLURM, i.e. basic, mode. Do not set anything for SLURM mode, the launcher will handle it. This is due to Vulkan/PyTorch differences in GPU indexing.

An example of training SpawnNet DAgger on the Open Drawer task is found in scripts/sim_exps/spawnnet_exp.sh.

1. Make sure to specify the ISAACGYM_ARG_STR as an environment variable (it should be the exact same value as the example).
2. For the drawer task, use isaacgym_task=open_drawer, and for the door task, use isaacgym_task=open_door_21.
3. Optional: Our framework splits 21 training assets among the allocated GPUs. Each asset has a corresponding simulation environment that's assigned to the same GPU as the asset. By default, each GPU gets floor(21 / num_gpus) assets (with the remainder assets going to the last GPU). If you wish to split the assets differently, set the variable TRAIN_ASSET_SPLIT as follows when kicking off the train.py script:

   TRAIN_ASSET_SPLIT=<# assets on 0th GPU>,<# assets on 1st GPU>,<# assets on 2nd GPU>,...

   When a larger model is being trained, the primary GPU (where the model resides) may run into CUDA memory issues from sharing space with too many simulation environments. The other GPUs may have space to load more environments. This fix is helpful for that case. Note that this custom asset splitting only applies for training.

We provide entrypoints for each experiment in scripts/sim_exps.

Debugging Training

We provide a script, scripts/sim_exps/sim_debug.sh, to assist with debugging training in simulation. This script enforces only one environment, one GPU, to be used.

You can run the script as is to test that the spawnnet simulation framework is functioning correctly. You can also test different methods, tasks, and seeds by following the comments in the script. Leave the ISAACGYM_ARG_STR as is, to ensure only one environment is loaded (for faster testing).

Note: This script always runs with only one GPU.

Evaluation

If you're in SLURM mode, evaluations are handled automatically by our training script and can be found under the eval folder of your run. Statistics will be listed under summary.csv.

Otherwise (basic mode), due to memory issues with IsaacGym, evaluations must be handled manually, on any single GPU. The syntax is the same regardless of the experiment done:

export ISAACGYM_ARG_STR="--headless --rl_device=cuda:0 --sim_device=cuda:0 --cp_device=cuda:0 --test --use_image_obs=True"
WHICH_GPUS=0 python RLAfford/dagger/eval.py <your exp dir> --chunk_id 0 --num_chunks 1 --mode basic

and results get saved the same way as SLURM.

Real World Experiments

For Real World BC Experiments, the demonstration set for the Place Bag task can be found at this Google Drive link. You can download this with gdown. After downloading, place the unzipped directory into /dataset.

Similarly to simulation, we provide entry points under scripts/real_exps.

Visualizing pre-trained feature attention

After running a SpawnNet experiment, visualizations of the adapter features can be found under the run's directory, which looks like:

/data/local/0627_place_bag_spawnnet_2050/0/visualization_best

Adding Tasks

To add tasks, please refer to simple_bc/constants.py, and follow the format for either BC_DATASET or ISAACGYM_TASK.

Acknowledgements

The gdict library is adopted from ManiSkill2-Learn. Additionally, we use tasks and assets from RLAfford.

THIS SOFTWARE AND/OR DATA WAS DEPOSITED IN THE BAIR OPEN RESEARCH COMMONS REPOSITORY ON Jul. 7., 2022.