LunaVLA v3.0.0-rc.1

Release candidate for the reproducible LunaVLA v3.0 teaching core.

This prerelease contains the Python package, SBOM, provenance, API and migration contracts, exact SHA256SUMS, a 1,550-row deterministic teaching-fixture evidence bundle, and the offline Portfolio bundle. The signed tag points to v3-next@b5a93faf7c28a8fba0500cd862410896799d52e3.

Scope is intentionally bounded: no model weights, checkpoints, caches, raw/real datasets, PyPI publication, physical-robot claim, or PushT/LIBERO benchmark claim.

LunaVLA v3.0.0-alpha.3 — CPU policy contracts

Code-only prerelease: ACT and Diffusion CPU paths plus a SmolVLA public-API conformance adapter. No pretrained weight, policy-performance, modality, task, or robot-deployment claim.

LunaVLA v2.0.0 — Stable CPU teaching and evidence release

LunaVLA v2.0.0 freezes and releases the reviewed v2 teaching/research bridge on signed source commit 72b052864229661e1d5e1ff7692966630b84d9af. The stable source tree is identical to protected v2@40ec92eb4a8e4340acd42e3566a7d31cea0b0933.

Authoritative post-merge gates:

• stable evidence + real LeRobot integration: https://github.com/xiaoms22/lunavla/actions/runs/29132868486
• Fast CI: https://github.com/xiaoms22/lunavla/actions/runs/29132857148
• V2 CPU: https://github.com/xiaoms22/lunavla/actions/runs/29132857151
• CodeQL: https://github.com/xiaoms22/lunavla/actions/runs/29132857168

Evidence scope:

• the canonical language and visual designs were rerun after merge for exactly 15 training runs and 960 arm-episodes;
• seed-11 language and image reruns reproduced checkpoint, configuration, data, and metrics hashes exactly;
• the pinned official LeRobot PushT episode and headless gym-pusht adapter passed on the same source SHA with verified GitHub/Sigstore provenance;
• the combined archive contains full outputs, review snapshots, frozen contracts, distributions, environment, SBOM, integration provenance, and internal checksums.

Claim boundary:

• Instruction-following has not yet been established.
• Visual-control contribution has not yet been established.
• The LeRobot integration proves adapter connectivity only; it is not a PushT performance claim.

CPU Linux is the authoritative supported evidence environment. CUDA remains an experimental manual path and is not claimed as verified. This release is not a real-robot, production deployment, or PushT benchmark release. No package is uploaded to PyPI.

Verification:

1. Download lunavla-v2.0.0-release-assets.tar.gz and lunavla-v2.0.0-release-assets.SHA256SUM.
2. Run sha256sum -c lunavla-v2.0.0-release-assets.SHA256SUM.
3. Extract the bundle, enter release-assets/, and run sha256sum -c SHA256SUMS.
4. Verify provenance with gh attestation verify SHA256SUMS --repo xiaoms22/lunavla.

The convenience bundle preserves the workflow artifact directory layout that GitHub otherwise flattens for individual release assets. The inner SHA256SUMS and release subjects retain the authoritative GitHub/Sigstore provenance from the stable workflow.

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.