Bridge Data Robot Infrastructure

Platform for data acquisition and training with WidowX in docker container. This is an accompanying code repository for the paper "Bridge Data: Boosting Generalization of Robotic Skills with Cross-Domain Datasets" (arXiv paper). Here is the project website where you can find more information about how to use and contribute to the dataset. All the imitation learning model training code using the dataset can be found at the bridge data imitation learning repository.

Project structure

'docker' folder - contains Dockerfile and docker-compose.yml file.
'widowx_envs' folder - docker container-mounted python package.

Environment variables

The following environment variables are used and can be adjusted to mount given directory in the docker container:

• CODE - mounted as '~/code' folder (defaults to 'code' folder in the repository root)
• DATA - mounted as '~/trainingdata' folder (defaults to 'trainingdata' folder in the repository root)
• EXP - mounted as '~/experiments' folder (defaults to 'experiments' folder in the repository root)

Important

In order to make sure the default values are used and required directories are created with user ownership keep using scripts/docker-compose proxy instead of plain docker-compose when running the container.

Host machine setup

In order to set up the host machine to run robonetv2 run:

bash scripts/host_install.sh

Pull image and run container

To avoid specifying the arm type in docker container after each container recreation, set up ROBONETV2_ARM environment variable.
For example:

echo "export ROBONETV2_ARM=wx250s" >> ~/.bashrc && source ~/.bashrc

Then run the container.

scripts/docker-compose \
   -f ./docker/docker-compose.yml \
   up \
   -d \
   --force-recreate

Build image and run container

In order to use the provided Dockerfile.base and rebuild the image with your changes use 'docker/docker-compose.build.yml' file:

scripts/docker-compose \
   -f ./docker/docker-compose.build.yml \
   up \
   -d \
   --force-recreate \
   --build

Access the container

docker exec -it robonetv2_${USER} /bin/bash

USER variable is used to avoid the clashes between the containers of different users.

Restart the container

scripts/docker-compose \
-f ./docker/docker-compose.yml \
up \
-d

Cleanup

scripts/docker-compose \
   -f ./docker/docker-compose.yml \
   down \
   --rmi all \
   --volumes

Run the code

Run the stack with one command

~/scripts/run.sh

For the available arguments of run.sh call ~/scripts/run.sh -h

To verify the robot functionality call the exemplary training environment:

python ~/widowx_envs/widowx_envs/widowx/widowx_env.py

After finishing the interaction before the shutdown you can call go_sleep from bash to move the robot to the sleep position.

Data collection with teleoperation:

cd ~/widowx_envs
python widowx_envs/run_data_collection.py experiments/toykitchen_fixed_cam/conf.py

You can change data collection parameters in experiments/toykitchen_fixed_cam/conf.py

Running an imitation learning policy

cd ~/widowx_envs
python widowx_envs/run_data_collection.py experiments/toy_kitchen_v0_control/conf_highres.py

You can change the policy rollout parameters in experiments/toykitchen_fixed_cam/conf.py

Global parameters for WidowX envs

Each WidowX environment is reading 'widowx_envs/global_config.yml' created at the first run of 'scripts/run.sh'. That's a file which can be used as set of parameters which will be used by all WidowX envs (including teleoperation). For example, this file can be used to rotate the workspace in all WidowX environments on the specific machine by adding this line:

workspace_rotation_angle_z: 1.57

This configuration change will be also used by the go_sleep function. In order to avoid the changes in the configuration file from showing in git, please call:

git update-index --assume-unchanged widowx_envs/global_config.yml

To revert it, call:

git update-index --no-assume-unchanged widowx_envs/global_config.yml

git update-index does not propagate with git, and each user will have to run it independently.

Using realsense cameras

The realsense cameras require different drivers than RGB cameras. If you are using realsenses, change the camera_string in scripts/run.sh to realsense:=true.

You will also need to update the device IDs in /widowx_envs/widowx_envs/widowx/launch/realsense.launch to match your cameras.

Troubleshooting

Permission errors

If you run into following errors:

Traceback (most recent call last):
  File "urllib3/connectionpool.py", line 677, in urlopen
  File "urllib3/connectionpool.py", line 392, in _make_request
  File "http/client.py", line 1277, in request
  File "http/client.py", line 1323, in _send_request
  File "http/client.py", line 1272, in endheaders
  File "http/client.py", line 1032, in _send_output
  File "http/client.py", line 972, in send
  File "docker/transport/unixconn.py", line 43, in connect
PermissionError: [Errno 13] Permission denied

that can be fixed by running the following commands and subsequently restarting the PC (the log out and log back in is sometimes not sufficient):

sudo groupadd docker
sudo usermod -aG docker $USER