Structure of the User Code Package

In order to be able to execute your code on our robots, you need to follow a few rules regarding the structure:

All code needs to be in a single git repository. You may use git submodules to add content of other repositories.
Your code should be provided as a regular Python package that can be installed by calling:
```
pip install .
```
at the root directory of the repository.
Apart from your code, the repository needs to contain a configuration file called trifinger.toml, see Configuration File.

Note

We provide an example package (the one that was already used for the pre-stage). You can use this package as base for your own one.

Python Package

When running your code on the robot, your git repository is expected to contain a Python package that will be installed via pip.

This means that if your code depends on other Python packages, that are not already included in our Apptainer image, you can simply specify them as dependencies (e.g. in the setup.cfg when using setuptools). If you have dependencies that cannot be installed with pip, you can provide a custom Apptainer image (see Add Custom Dependencies to the Container).

Configuration File

Your git repository must have a TOML file trifinger.toml in its root directory which specifies the classes that implement your policies as well as the task you want to run. For example if your package is called my_trifinger_rl_package and contains a module policies in which the classes MyPushPolicy and MyLiftPolicy are defined:

[trifinger_rl_experiment]
task = "push"  # "push" or "lift"

# Specify a policy for each task.
push_policy = "my_trifinger_rl_example.policies.TorchPushPolicy"
lift_policy = "my_trifinger_rl_example.policies.TorchLiftPolicy"

task has to be set to either “push” or “lift” and determines which environment is used when you submit a job to the robot.

{TASK}_policy specifies the class that implements the policy for the corresponding task.

To run jobs on the robots, it is enough if you specify a policy for the task set in the file.

The Policy Classes

The policy classes specified in the configuration file (see above) are expected to implement the interface of trifinger_rl_datasets.PolicyBase. See Starting from the Example Package for more information.

Run Code inside the Container

To locally test your code inside the Apptainer container, please follow the instructions given in the simulation section: Set up Evaluation Environment using Apptainer.

Validate Your Package Before Submitting to the Robots

The example package contains a script validate.py which runs some basic checks to verify that your package has the correct structure (e.g. if the package can be installed, if a policy is configured and if it follows the expected interface, etc.). It is recommended to run this script to check for basic errors before you start submitting jobs to the robots. Assuming your package is based on the example package, you can run it like this from the root directory of the package:

cd path/to/your_package
apptainer run -e --no-home --bind=$(pwd),${HOME}/.cache/pip ./path/to/trifinger_rl.sif python3 scripts/validate.py

Replace trifinger_rl.sif accordingly, if using a custom container.

Binding ~/.cache/pip is not strictly needed but it avoids unnecessary downloads of packages which can make a significant difference if large packages (e.g. torch) are involved.