---
pipeline_tag: robotics
library_name: microduck
tags:
- microduck
- onnx
- tug-of-war
---
# dfreq-tug — declared-start specialist 160168

`policy.onnx` is an attributed derivative of our frozen T19 policy with an **explicit benchmark-specific observation lookup**. Local official-runner mean on declared starts 1–5: **160.1679945 cm**, with five pit wins. This is **in-sample**: those starts were used for optimization. It is not an Arena submission receipt, leaderboard/rank claim, or evidence of generalization.

Model SHA256: `3d67cd3705bf8c8e12ee27723edc05b378fe8c7274350d61c7e53f19b8cfc362` (974,598 bytes).

## Policy behavior
The stateless ONNX accepts the unchanged legal raw 61-observation vector and produces 14 position offsets (`DEFAULT_POSE + action`). It matches 515 observations from our own optimized trajectories with coordinatewise/L-infinity tolerance 1e−4. Matching regions are non-overlapping in the frozen table. A match returns its recorded optimized action; unmatched observations use **exact T19 fallback**. Lookup uses all legal observation fields, including previous actions; no seed, scoring clock, world position, opponent state or privileged critic input is appended.

The stored trajectories were optimized for the declared starts. A legal-observation lookup does not make this a robust learned feedback controller: small deviations can leave the table and invoke the weaker fallback. No robustness claim is made. Exact fallback on consumed starts verifies unchanged fallback behavior, **not improved performance on unseen starts**.

## Local validation
Official Space revision: `76dc0e06a6d261bcf9bdf7cce3fdc062c3c14b23`. The current core runner/actuator/Tug/constants/event files were checked against this local revision; hashes are in `manifest.json`. Official physics, preparation, delays, opponent and scoring were not modified.

| Declared start | Score(cm) | Outcome |
|---|---:|---|
|1|161.71196699|pit win|
|2|159.91537571|pit win|
|3|158.68700147|pit win|
|4|159.58207250|pit win|
|5|160.94355583|pit win|

All five exported trajectories replayed exactly against their frozen teachers. On 96 already-consumed development starts (176001–176032,177001–177032,178001–178032), the export reproduced T19 fallback trajectories exactly. **Zero fresh reserve starts were accessed for this validation.** 197 native validation bouts comprise five specialist runs plus 96 paired specialist/T19 comparisons. An additional pure 100-unmatched-observation check reproduced T19 outputs exactly.

Maximum recorded rope-end tie gap across the five starts/preparation was 5.2964 mm, below the 10 mm conservative local guard and 50 mm official fault threshold. Available recorded frames are checked; this is not a claim to have exported every internal solver substep. `validation-visual.png` and `visual-review.json` document inspected collision-geometry reconstructions; these do not measure contact forces.

## Method, provenance and rights
This is an **attributed official-starter neural remix and own-trajectory specialization**, not a from-scratch policy. Base: [djduecaster/dfreq-tug T19](https://huggingface.co/djduecaster/dfreq-tug/tree/6bb0a485a2d27c4a5511d58fc92ca781ad604c77), model SHA256 `61707f5f3aca15b0a5bec1ff68ae119bbc840f6fdc3d618953791ab52b7cb6d1`.

As documented by that pinned T19 card, the base inherits **Nico-robot's `mjlab-tugofwar-microduck-t20` checkpoint 500** neural backbone and baked normalizer through the officially provided **Pollen Tug starter/reference**, SHA256 `245869c02049f29f453d48619969b2ef51d73d9751813ddac2b979be6a3b2e44`. T19 preserves the original backbone/normalizer and adds own observation-feedback adapters/internal request offsets. This specialist preserves that T19 branch as fallback and adds own optimized trajectory observations/actions. Our development included CPU neural-policy training and derivative-free trajectory search; the final lookup is not itself a newly trained general-purpose neural controller.

No leader policy/code or unlicensed stronger checkpoint was copied into this specialization. **No new standalone redistribution license is asserted**; see `LICENSE-NOTE.txt`. Upstream rights remain applicable. Do not infer a permissive license from repository accessibility.

## Files
`policy.onnx`, `manifest.json`, `validation.json`, `visual-review.json`, `validation-visual.png`, `LICENSE-NOTE.txt`, and `SHA256SUMS`. This compact package excludes training corpus, simulator assets and separate third-party weights/source. The model necessarily retains its attributed inherited T19 weights. Source hashes document lineage; this is not a full training-reproduction bundle.