SkiREIL - Skill-based Multi-objective Reinforcement Learning of Industrial Robot Tasks with Planning and Knowledge Integration

Matthias Mayr Faseeh Ahmad Konstantinos Chatzilygeroudis Luigi Nardi Volker Krüger

Paper on arXiv - Code on Github

We present SkiREIL, an approach to learn (single- and) multi-objective tasks by utilizing existing skills that expose learnable parameters. Targeted at industrial environments where tasks often have several objectives (KPIs) such as task performance and safety, it provides a way to retrieve a large variety of solutions from which the robot operator can choose from. By integrating planning and a knowledge framework it can offer a pipeline from high-level goal definition to the execution of learned policies on a real system while conducting learning itself in highly parallelized simulations.

Implementation and Configuration of Scenarios

SkiREIL is intended to be a flexible and configurable approach that can support a variety of scenes, robots and learning strategies.

We show that SkiREIL can learn contact-rich tasks entirely in simulation and outperform robot operators parameterizing the same skills:

However, more scenarios are already available and show equally good performance at learning a variety of strategies that translate well to a real system.
By utilizing robot_dart, the definition of new scenes and robots is easy. Common formats such as URDF are supported and important utility functions are readily available.

Source Code

The source code is available here: https://github.com/matthias-mayr/SkiREIL
Feel free to open issues or send pull requests