Open source Active Apache-2.0

Robot Skills

An open, platform-independent format for packaging humanoid robot skills, like opening a bottle or ironing a shirt, as compact, signed, nestable packets that can be tested in simulation and shared through an open library.

Overview

A robot skill should be something you can hand from one robot to another, like a recipe. Robot Skills (RS) is an open draft format that describes what to do and where things move relative to objects, rather than the joint angles of one particular robot. Each robot maps a skill onto its own body.

Status: early draft (v0.1), open for discussion. The repository has the specification, examples, a PostgreSQL library prototype and a Python validator for YAML/JSON skill sources. There is no packet compiler or robot runtime yet, and no example is an executable skill.

Principles

  • Portable: skills describe goals and object-relative motion, not one robot’s joints.
  • Compact: a typical packet is a few kilobytes; motions are small mathematical primitives, not recordings.
  • Nestable: “iron a shirt” can reuse “grasp an object” and “lay a garment flat”.
  • Learnable: a robot downloads a skill, adapts it to its own tools and can publish improvements as variants.
  • Trustworthy: every skill carries provenance (author, origin, tests) and is signed by its author and, once tested, by the library.
  • Open: skills can be written by hand or recorded from a person, tested in simulation such as MuJoCo, and shared through an open library.

Safety

Robot safety remains the robot manufacturer’s responsibility. RS describes skills; it does not replace manufacturer safety controls or certify safe execution.

Get involved

Ideas, reviews and criticism are welcome as GitHub issues. See the overview for an example skill and the main design choices.

Licence

The specification, examples and tools are Apache 2.0. Skills published to the RS library default to CC BY 4.0 (authors may choose CC0).