Four test scenarios in RViz (Kinova Gen3).
A versioned, checkable package format for robot skills. The .whl and pip check for one-tap skills, not the skill and not the marketplace.
An August 2026 survey of commercial robot-skill marketplaces notes that they already ship one-tap skills, but only as static playback. That leaves six properties open: adaptation, cross-embodiment portability, provenance, safety verification, composition, and standardisation. This repo does not build a marketplace or a learning system. It ships a concrete manifest format plus reference tooling that makes those six properties explicit in the package and checkable by a command.
One skill.yaml manifest carries all six properties a static-playback skill leaves implicit.
- A strict, versioned
skill.yamlschema (standardisation) with a validator that names the offending field. - A content id and per-payload sha256 integrity map (provenance) so a skill has a stable identity and tampering is caught.
- A declared adaptation contract and an
adapt()that remaps a skill onto another arm or refuses and names the missing mapping (portability, not silent playback). - A safety block with force/speed/clearance limits and a safe precondition/postcondition language, and a verifier that refuses a skill whose limits exceed the target robot's or whose preconditions are false.
- A composition block whose sub-skill handoffs are checked: step N's postconditions must establish step N+1's preconditions.
- A reference CLI and library that run all of the above today, with no robot.
It is not a marketplace, not a policy or a learner, and not a runtime. It carries no images or model weights. It is the envelope and the checks; the skill inside is yours.
pip install git+https://github.com/megazron/skillcrate-spec
Or clone and run from source with PYTHONPATH=src.
skillpkg init my_skill # scaffold a manifest
skillpkg validate my_skill # schema + semver
skillpkg verify my_skill # recompute payload hashes
skillpkg portability my_skill --target other_arm.yaml
skillpkg adapt my_skill --target other_arm.yaml -o my_skill_ported
skillpkg compose pick_and_place --registry ./registry
skillpkg pack my_skill -o my_skill.skill
skillpkg selftest
The bundled example walks all six properties end to end, including the deliberate failures:
python3 examples/run_example.py
Every check is a gate. A skill that fails one is refused with a reason; it never runs on best effort.
schema_version: 1
name: pick_cube
version: 1.0.0 # semver, validated
description: Pick a cube from the table with a parallel gripper.
provenance: # where it came from, and its lineage
author: megazron
created: '2026-08-21'
source: recorded # recorded | learned | authored | composed
captured_on: gen3_left # must match the embodiment below
episodes:
- {id: ep_2026_08_21_014, sha256: 0000...}
derived_from: [] # parent skill ids
embodiment: # the arm it was captured on
name: gen3_left
dof: 7
joint_names: [joint_1, ..., joint_7]
ee_frame: tool_frame
gripper: robotiq_2f85
control_mode: position
units: {angle: rad, length: m}
adaptation: # how to run it elsewhere -- a contract
joint_map: {} # source_joint -> target_joint
frame_transform: {} # e.g. {ee_frame: tool0}
unit_conversion: {} # e.g. {angle: rad->deg}
requires: [] # remaps that MUST be provided to port
safety:
max_force_n: 20.0
max_speed_mps: 0.25
min_clearance_m: 0.12
preconditions: [gripper_open and cube_visible and table_clearance_m > 0.12]
postconditions: [cube_grasped]
verified: {tool: skillpkg, when: '2026-08-22', result: pass}
payload: [trajectory.csv] # the actual recorded motion
integrity: {} # sha256 per payload, filled on saveStandardisation. skillpkg validate fails on a bad semver, a missing provenance field, a joint_names list that does not match dof, or a precondition that does not parse. A pull request that does not validate fails CI.
Provenance and integrity. The package id is a content hash over the manifest plus the payload integrity map, so the same skill always hashes the same way and any change moves it. skillpkg verify recomputes every payload sha256 and fails on a mismatch. provenance.derived_from records the parent ids, so lineage is walkable.
Portability. Portability is a declared contract, not a hope. skillpkg portability my_skill --target other.yaml refuses if the target's joints, units or end-effector frame differ and the manifest does not declare the remap. Fill in adaptation and it passes. This is the concrete line between static playback and a portable skill.
Adaptation. skillpkg adapt applies the declared joint map, frame transform and unit conversion, writes a new package with a bumped version and the parent id in derived_from, or refuses and names the missing mapping. It never guesses.
Safety verification. skillpkg verify-safety --robot robot.yaml --world world.yaml refuses a skill that assumes a force, speed or clearance the target robot cannot honour, and refuses to green-light one whose preconditions are false in the given world. Conditions are evaluated by a whitelisted ast parser: no calls, no attribute access, no imports.
Composition. A composed skill lists sub-skills by reference. skillpkg compose --registry ./registry resolves each and checks the handoff: step N's postconditions must include every precondition step N+1 needs. A gap is a failure with the exact missing condition, so composed skills are validated, not just concatenated.
The format and checks are real and runnable; the hard parts they make checkable are not thereby solved. Adaptation applies declared remaps but does not learn one; the safety verifier checks declared limits and symbolic preconditions, not dynamics; composition checks symbolic handoffs, not physical feasibility. It is the contract and the gate, which is what was missing.
MIT, see LICENSE.