kevinqz/Robometer-4B-CoreAI

Canonical: kevinqz/Robometer-4B-CoreAI — source of truth.

Robometer-4B Reward (fabric)

An Apple Core AI conversion of lerobot/Robometer-4B — the deployable reward-head core of a robot-policy reward model. It maps per-frame vision-language hidden states to progress (a distribution over discrete bins) and success logits, for reward/progress estimation in robot learning. Produced by coreai-fabric and indexed by coreai-catalog.

Reward heads, not the whole model — this needs the VLM backbone you supply. Following the split discipline of the VLA lanes (EVO1 / VLA-JEPA / pi0), this asset ships ONLY the small MLP reward heads. The host owns the Qwen3-VL backbone (a standard VLM), the <|prog_token|> hidden-state extraction, and the decode (progress = softmax-weighted bin-mean clamped to [0,1]; success = sigmoid). Without the backbone + processor the graph is inert. This is a conversion-fidelity artifact, not a benchmarked reward signal.

Model facts

Field	Value
Parameters (full model)	4.45B
Architecture	transformer
Capabilities	reward-modeling, robotics
Hidden dim (VLM)	2560
Progress bins	10
Max frames (static)	8
Outputs	progress_logits, success_logits
Quantization / precision	none / float32
On-disk size	25 MB
Asset kind	MLP reward heads (VLM hidden states -> progress + success logits)
assetVersion	2.0

Use it — this needs host code you supply

The bundle is a single static graph: per-frame hidden states frame_embeddings [1, T, hidden] in → progress_logits [1, T, bins] + success_logits [1, T] out. You supply the Qwen3-VL backbone that produces those hidden states at the <|prog_token|> positions, plus the decode, in your host code (Swift or Python). Use the upstream repo for the backbone + processor.

pip install coreai-catalog && coreai-catalog install robometer-4b

Requirements

• Deployment: macOS 27.0+ / iOS 27.0+, Xcode 27+. The asset serializes with minimum_os v27, so the on-device Swift runtime requires macOS/iOS 27+. A Mac on macOS 26 can convert and inspect it but not run it on-device.
• Apple Silicon.
• The upstream Qwen3-VL backbone + Robometer processor (host-side) to produce the input hidden states.

Verification (output parity)

• Gate A (structure): passed — the bundle's layout + metadata were validated; the graph loads.
• Gate B — graph_output_cosine: 1.000000 min output cosine (median 1.000000) vs the fp32 torch reward heads over 8 seeded hidden-state inputs (worst of the progress + success heads), measured on apple_silicon. Certifies the export computes the SAME reward-head logits as the source — a conversion-fidelity metric, not reward quality.
• This certifies the export is numerically faithful to the source reward heads — it does NOT certify reward quality or downstream task success. Reproduce with coreai-fabric verify.

Provenance

Field	Value
Base model	lerobot/Robometer-4B @ db167a7c369a3ee59cda801fe33ca9da560b1662
Converted by	models/robometer/export.py (version not reported)
Recipe	robometer-4b (recipe_source: fabric)
Precision / quantization	float32 / none
Conversion date	2026-07-07

Machine-readable, in this repo: parity-report.json · reproduce-manifest.json · LICENSE.

License and attribution

Weights licensed apache-2.0 — see the bundled LICENSE. This artifact is a converted derivative of the base model's reward heads: their weights were converted to Apple Core AI format. The conversion itself is community work.

Links

• Base model: lerobot/Robometer-4B
• Reproduce: recipe robometer-4b
• Index: coreai-catalog
• HF Collection

The on-device Core AI ecosystem

• coreai-fabric — the reproducible recipe → .aimodel pipeline that produced this asset.
• coreai-catalog — the index of Core AI models with provenance and integration snippets.
• apple/coreai-models — Apple's official exporters and runtimes.

Not affiliated with Apple

Community conversion. Not produced, hosted, or endorsed by Apple. Apple and Core AI are trademarks of Apple Inc., used here only to describe the target runtime/format.