GETok-DSR: Vocabulary-Grounded Draft-Select-Refine Planning
This repository contains the frozen NAVSIM v1 submission and the minimal code needed to reproduce its candidate-aligned progressive GETok selection over an immutable ChainFlow-VLA fallback.
Method
The released ChainFlow-VLA trajectory is retained as candidate alpha 0. A
route-conditioned GETok/DrivoR draft supplies alpha 1; circular heading
interpolation and linear position interpolation produce alpha
0.25/0.50/0.75. A navtrain-only five-fold stability screen caps deployed
support at alpha 0.50. The remaining fallback/alpha025/alpha050 candidates
are each passed through the same frozen visual DrivoR scorer, and the candidate
with the highest predicted composite score is returned:
fallback -> alpha025 -> alpha050 | alpha075 -> GETok draft
deployable support | diagnostic support
frozen candidate-aligned DrivoR score -> argmax
The final selector retains the fallback for 1,967 of 12,146 navtest scenes and
chooses alpha 0.25/0.50 for 1032/9147 scenes, respectively.
Reproduction
Install the official NAVSIM v1.1 environment and ChainFlow-VLA release first:
- NAVSIM v1.1: https://github.com/autonomousvision/navsim/tree/v1.1
- ChainFlow-VLA: https://github.com/AFARI-Research/ChainFlow-VLA
The scripts under code/ include the original GETok support construction and
the current progressive-vocabulary builder and candidate-aligned submission
packager. They expect the official NAVSIM v1.1 dataclasses, the released
ChainFlow-VLA fallback, and DrivoR proposal-score payloads.
python code/apply_chainflow_paired_consensus.py --help
python code/build_progressive_vocabulary.py --help
python code/package_candidate_aligned_submission.py --help
python code/package_capped_direct_submission.py --help
python code/finalize_navsim_submission_metadata.py --help
The uploaded submission.pkl is the exact rank-1 official submission. Its
trajectory digest and source-submission SHA-256 are recorded in
artifacts/g339_official_manifest.json; the navtest score is stored in
artifacts/g339_navtest_summary.json, and the navtrain-only policy search in
artifacts/g338_policy_search.json. The preceding G247/G332 submissions are
retained under artifacts/ for provenance.
Before publishing, verify the immutable package:
python verify_release.py
With a write-scoped Hugging Face token installed, publication is one command:
HF_ENDPOINT=https://huggingface.co python publish_hf.py
Evaluation Boundary
The local official-protocol NAVSIM score is 0.954545187187109, compared with
0.9491968721563916 for the exact released ChainFlow-VLA submission on the
same 12,146-scene cache. The local score is above the public 0.953087
reference checked on 2026-09-08. The preceding uncapped G332 submission
received official public PDMS 95.2814871298245. G339 received official public
PDMS 95.36882969298688, ranking first and exceeding DriveZero 95.3087 by
0.06012969298687665 points. The official component scores are NOC
99.19726658982381, DAC 99.56364235139141, DDC 97.21307426313189, EP
92.51793711640666, TTC 97.29952247653549, and comfort 100.0.
The evaluated immutable repository revision is
403c00322c64882ba6601669c40b52194ab57012; the structured server return is
stored in artifacts/g339_official_result.json.
Authors
YuHuan, Wuhan University (WHU). Contact: yuhuanzjj@whu.edu.cn.
Li Ming, Wuhan University (WHU). Contact: Liming751218@whu.edu.cn.
Yuanyuan Shi, Wuhan Technical University. Contact: yuanyuanshi.research@gmail.com.
ZhangPin, Wuhan University (WHU). Contact: pzhang1992@whu.edu.cn.
The initial release namespace is minfishmovemountain. The repository may be
transferred to the planned WHU-LiFish-Lab organization after it is created;
the method name and frozen trajectory payload will remain unchanged.
Acknowledgements
This method uses the released ChainFlow-VLA checkpoint and submission as its strong immutable fallback. NAVSIM and ChainFlow-VLA remain governed by their respective licenses and citation requirements.