KRKnown RobotALPHA
Reproduction sprints

OPEN UTILITY · VERSION 1.2.0 RELEASE CANDIDATE

Validate robot-policy reproducibility.

Check a LeRobot policy repository before attempting reproduction on an SO-100 or SO-101 arm. The validator inventories reproducibility evidence, writes a portable robot-skill.yaml, and identifies missing metadata. Configuration comparison does not prove physical transfer or safe operation.

policy / terminal
$ robot-skill check ./policy --strict

INCOMPLETE  ./policy
Detected 4 evidence files
ERROR  policy.framework_version
  Pin the framework version.
ERROR  hardware.gripper
  Declare the end effector.
WARN   evaluations
  Compatibility remains unverified.

01 / QUICKSTART

From repository to manifest

  1. 1
    Installpipx install . # from the Known Robot 1.2.0 checkout
  2. 2
    Inspect a policyrobot-skill check ./policy
  3. 3
    Use it in CIrobot-skill check . --strict --no-write

Python 3.10–3.12 · macOS, Linux, and Windows · Apache-2.0

02 / EVIDENCE CONTRACT

One file. Eight evidence groups.

The manifest stays beside the policy on GitHub or Hugging Face. Null values are allowed structurally so the first scan can be saved; strict mode fails until every required claim is supplied.

01Policy

Framework, pinned version, architecture, checkpoint

02Hardware

Robot family, gripper, sensors, calibration references

03Runtime

Control frequency, observation and action shapes

04Dataset

Repository, revision, feature schema, license

05Dependencies

Resolved package versions and hashed declarations

06Source

Repository type, immutable revision, artifact version

07Compatibility

Supported, unsupported, and untested hardware

08Evaluation

Benchmark, trials, successes, evaluator, evidence

03 / RELIABLE AUTOMATION

Designed to fail usefully.

Exit 0

The scan ran. In default mode, an incomplete draft is still written so contributors can improve it.

Exit 1

The path, manifest, or filesystem could not be read. No misleading output is written.

Exit 2

Invalid evidence always fails. Strict or complete validation also fails on missing metadata; target comparison fails on mismatches or unknowns.

JSON diagnostics

--format json returns stable finding codes, paths, severities, and a summary for CI.

Use --target ./target.yaml --no-write --format json to compare declared configurations. Equal shapes alone are not sufficient: feature semantics and calibration matter. A match never establishes safe physical transfer or independent reproduction.

STANDARDIZE IN THE OPEN

Use the format without joining a platform.

The schema, example, implementation, and tests are public. Keep the manifest with your policy and review it like code.

Read the CLI guide ↗Download schema 1.0 ↗Reusable GitHub Action and portable credit guide ↗Inspect adapter ownership →Corrections and disputes →Report a transfer failure →