LunaVLA v1.1.0 — Trustworthy teaching core

LunaVLA v1.1.0 — trustworthy teaching core

This release resets LunaVLA around a claim boundary that matches the implementation: a CPU-runnable, state-based NumPy imitation-learning/vision-action teaching loop. It is not an ACT Transformer, a real PushT benchmark, a visual VLA, or a robot-deployment stack.

Highlights

• Canonical numpy_linear_chunk and numpy_bc_mlp policies.
• Versioned ExperimentConfig, ActionChunk, VLARecord, checkpoint, and RunManifest contracts.
• Correct masked MSE gradient and padding behavior.
• Episode-disjoint train/validation/test splits and strict JSONL validation.
• Receding-horizon and open-loop-chunk evaluation.
• Python 3.10–3.12 CI, 87 tests, 90.14% core coverage, CodeQL, Dependabot, and full-history secret scanning.
• Controlled 5-seed × 20-evaluation-episode experiments with Wilson and paired-bootstrap intervals.

Controlled result boundary

The controlled chunk sweep does not support the historical claim that action chunking improves this task. Chunk size 1 performed best in the v1.1 experiment. The README result table is generated from validated manifests and aggregate summaries.

Verification

• Fast CI and CodeQL: https://github.com/xiaoms22/lunavla/actions/runs/29084214333
• Full Evidence: https://github.com/xiaoms22/lunavla/actions/runs/29084228841
• The annotated v1.1.0 tag is SSH-signed.

Download the evidence archive, SBOM, environment requirements, and SHA256SUMS below. Run sha256sum -c SHA256SUMS from the download directory to verify all assets.