talos-manipulation

Set of tools to work on manipulation tasks with the robot Talos.

This repo is built around several packages:

• deburring_deep_planner: a rosnode to do the inference of a neural-network
• deburring_mpc: an implementation of a Model Predictive Controller based on Crocoddyl
• deburring_python_utils: a collection of useful python tools
• deburring_ros_interface: the interface to allow the use of the MPC within a ROS architecture

Usage

The Apptainer is designed to provide a solution for running trainings of the RL policy or benchmarks. It is not intended for use when conducting experiments on the robot, thus it does not contain the deburring_ros_interface.

Build Apptainer:

apptainer build rl.sif apptainer/rl_mamba.def

Several apps are available:

• default training:

Short training with default parameters.

apptainer run --app default_training_mpc rl.sif

• training:

Training with custom parameters. This training requires a configuration file to be mounted in the Apptainer:

apptainer run --app training_mpc \
--bind path/to/config:/config \
rl.sif

An example of the config file to provide can be found in deburring_python_utils/gym_talos/config/config_MPC_RL.yaml.

• benchmark:

apptainer run --app benchmark rl.sif

The benchmark is carried out the configuration and the example policy that can be found in deburring_python_utils/deburring_benchmark/.

Copyrights and License

This package is distributed under a BSD-2-Clause Licence.

This repo was initially a fork of sobec.

Authored by:

• Alexis Delgorgue
• Côme Perrot
• Guilhem Saurel
• Olivier Stasse