starRoboHarness

An open execution and evaluation harness connecting robot policies, multimodal reasoners, and environments.

Observation → Policy proposal → Decision → Guarded execution → Trace

The core is model-neutral. QwenPI_v3, a Codex reasoner, and RoboDojo are reference integrations. Weights and simulator installations remain external; live integrations are experimental.

Quick start

git clone https://github.com/starVLA/starRoboHarness.git
cd starRoboHarness
bash scripts/bootstrap.sh

This installs the lightweight harness and runs a credential-free contract smoke test. It does not install RoboDojo/Isaac or download weights.

For real evaluations, configure your separate policy and simulator runtimes using deployment. Run preflight before launching a frozen panel:

.venv/bin/python -m starharness.rollout.local_campaign \
  --config local/runtime.json --panel configs/evaluation/robodojo_1task_v1.json \
  --output artifacts/preflight-001 --preflight

Replaceable components

Component	Responsibility
src/starharness/adapters	Policy proposal contracts
src/starharness/reasoners	Reasoner transport
src/starharness/rollout	Service lifecycle, execution, receipts and recovery
skills	Bounded robot tools and context
UI	Read-only monitoring

Python imports remain starharness for compatibility; the project and distribution are starRoboHarness.

Collaborate

Contributions are welcome. See contributing, roadmap, architecture, and evaluation.

Only complete native receipts contribute to SR/Score; infrastructure failures stay separate. Never silently replay a case after control may have started. The paired results include regressions and do not establish universal hybrid superiority.

MIT; see third-party notices.