Quchaosheng/workbench-mobile-home-robot

Verify before you say done.

An evidence-first foundation for mobile domestic robots: bounded actions, replayable events, and a verifier that can say confirmed, refuted, or insufficient evidence.

Product name: VORA Home Robot
Repository/runtime: workbench-mobile-home-robot (Workbench)

VORA is the product brand; Workbench remains the engineering repository and evidence-first runtime name. The VORA mark uses an open angle and off-axis orbit to stay distinctive without binding the identity to one robot use case.

简体中文 · Interactive 3D view

Robot demos often treat “command accepted” as “task complete”. Workbench keeps the proof in the loop:

goal -> bounded planner -> semantic action -> trusted executor
                                      \-> event store -> verifier -> replay/dashboard

Requirements: Python 3.12. No GPU is required for the offline runtime.

git clone https://github.com/Quchaosheng/workbench-mobile-home-robot.git
cd workbench-mobile-home-robot
python -m pip install -e ".[dev]"
python tools/scripts/sim_cli.py doctor
python tools/scripts/sim_cli.py run normal-001 --runner scripted --output-dir runs/demo

Run the full portable checks:

python -m pytest -q
python -m ruff check .

The scripted runner writes a replayable artifact (manifest, scene, events, logs, metadata, and SHA-256 checksums) labelled SCRIPTED_FIXTURE.

• Evidence-first verification with confirmed, refuted, and insufficient_evidence.
• Strict JSON schemas and matching Pydantic contracts.
• Append-only SQLite event storage with integrity-checked replay.
• Fail-closed policy validation for bounded semantic tools.
• Read-only dashboard and deterministic simulation fixtures.
• Software-only foundations for MCU, CAN, Motion, and BSP boundaries.

OmniLink AI is an independent service for searching maintenance notes, ADRs, issues, and Workbench run summaries. It is not a planner, executor, verifier, or robot-control dependency.

git clone https://github.com/vivekmaru/omnilink-ai.git
cd omnilink-ai
npm install
npm run dev                 # normally http://127.0.0.1:3000

Workbench uses the standard-library adapter in integrations/omnilink/:

from integrations.omnilink import OmniLinkClient

client = OmniLinkClient("http://127.0.0.1:3000")
results = client.search("gripper calibration")
answer = client.ask("Which calibration notes mention the gripper?")

Only bounded run summaries are exportable. Raw JSONL, TaskGraph, SemanticAction, action results, camera data, and safety state stay in Workbench. Catch OmniLinkError so an unavailable knowledge service cannot block offline operation. See the integration guide for deployment and security requirements.

The offline runtime and software boundaries are tested. End-to-end Gazebo worlds, real perception, semantic motion execution, and physical hardware evidence are not yet release claims. See docs/architecture/ and sim/README.md for the current evidence boundary.

• User guide
• Product evidence layer
• System architecture
• Simulation boundary
• Motion foundation
• Safety MCU
• Robot BSP
• Deployment
• Security policy

Generated from docs/evaluation/readiness-report-v1.json by tools/scripts/readiness_report.py. It states what the committed software is ready for; it grants no release approval and is not physical or Gazebo evidence. See multi-scenario readiness.

Apache-2.0. See LICENSE.

Generated from docs/releases/phase-status-v1.json by tools/scripts/phase_gates.py. It states which multi-scenario delivery phase proves itself; it grants no release approval and is not physical or Gazebo evidence. See phase gates.