An auditable, model-neutral harness for reliable and efficient robot-policy execution.
🤗 Hugging Face · 🌐 Homepage · 🇨🇳 中文文档 · 📄 Paper
starRoboHarness connects a robot policy, a reasoner, and an environment through one explicit control contract:
Observation → Proposal → Evidence gate → Decision → One controller → Native receipt
The harness keeps policy adapters, reasoner transports, simulator integrations, execution guards, and evaluation receipts separate. A policy can stay frozen, a reasoner can be replaced, and the native environment remains the authority for success.
- 📢 [2026/10/11] Released the starRoboHarness technical paper and the public release structure.
- 🔥 [2026/09/26] Added broader benchmark contracts and a more stable evaluation pipeline.
- 🔥 [2026/09/19] Open-sourced starRoboHarness.
The reasoner is a bounded decision-maker. A proposal is checked for freshness and action bounds before execution. The environment adapter owns episode progression, the controller records acknowledgements, and the native evaluator supplies the terminal outcome. The editable framework source is framework.pptx; the publication export is framework.pdf.
git clone https://github.com/starVLA/starRoboHarness.git
cd starRoboHarness
bash scripts/bootstrap.sh
python -m pip install -e '.[dev]'
pytest -qThe default installation contains the lightweight contracts and tests. Heavy simulators, checkpoints, and provider SDKs stay behind optional adapters. Read docs/deployment.md before launching an environment.
starRoboHarness/
├── src/starharness/ # stable contracts, adapters, reasoners, rollout
├── configs/ # versioned evaluation and runtime manifests
├── examples/ # small, runnable integration examples
├── scripts/ # preflight, launch, audit, queue, and reporting tools
├── tests/ # contract and integration tests
├── docs/ # architecture, protocols, benchmark, and operations
├── assets/ # logos, figures, paper, and release assets
├── UI/ # read-only local monitor and dashboard
└── benchmarks/ # benchmark contracts and result index
Simulation and real-robot integrations stay behind adapters. New integrations should add a backend-specific module and a contract test rather than importing simulator code into the core package. See docs/repository_layout.md and docs/design_principles.md.
A minimal contract loop is in examples/minimal_loop.py. To start the local read-only dashboard over a campaign directory:
./UI/run-dev.sh /path/to/campaign
# then open http://127.0.0.1:8766The dashboard reads evidence and serves task progress, camera frames, public decision records, and native receipts. It does not launch, stop, replay, or edit an evaluation. See docs/live_dashboard.md for remote port forwarding.
The figure below is the archived RoboDojo development comparison used by the technical report. It shows task-level native scores and completion counts for matched policy and hybrid configurations. It is included as a visual result record; the final paper and evidence ledger define the exact panel, denominator, and exclusions.
The editable source is robodojo-results.svg. The complete archived 50-case scene index from Table 12 is shown below and is also available in the benchmark results guide.
Raw rollout evidence and private operational logs remain outside the public repository.
The public tree records the contracts and entry points that are safe to reproduce. Current adapters cover RoboDojo and include a fail-closed VLAct contract check. Development integrations for RoboTwin 2.0, LIBERO-Pro, RoboCasa 365, BEHAVIOR, and SimplerEnv are documented with their scope and limitations in docs/benchmark_results.md and docs/vlact.md. A diagnostic or simulator smoke result is never presented as a full benchmark score.
The core package is hardware-neutral. Real-robot drivers belong in adapters that implement observation, action, safety-stop, clock, and native-receipt contracts. The public release documents the interface and current support boundary in docs/real_robot_support.md; it does not claim a physical-robot score without a native receipt.
The technical paper is available as starRoboHarness_tech_report.pdf.
If you find our ideas or environments helpful, please cite our work:
@article{starroboharness2026,
title = {starRoboHarness: A Community-Driven Efficiency Harness Infra for Uniting Robot Policy},
author = {{starRoboHarness Team}},
journal = {Technical report},
year = {2026},
url = {https://github.com/starVLA/starRoboHarness}
}We build on the open-source work of StarVLA, RoboDojo, RoboTwin, LIBERO, RoboCasa, robosuite, MuJoCo, and OpenPI. We also thank the broader robot-learning community for maintaining the environments, policies, and evaluation interfaces used by this project.
Contributions are welcome for adapters, evaluation protocols, diagnostics, documentation, and dashboard improvements. Start with CONTRIBUTING.md. Every benchmark change must state its protocol, denominator, seed, checkpoint identity, commit, and evidence location. Do not commit credentials, checkpoints, datasets, raw trajectories, or cluster paths.
MIT. Third-party components and their licenses are listed in THIRD_PARTY_NOTICES.md.