GETok-DSR: Vocabulary-Grounded Draft-Select-Refine Planning

This repository contains the frozen NAVSIM v1 submission and the minimal code needed to reproduce its GETok paired-consensus selection and immutable ChainFlow-VLA fallback overlay.

Method

The released ChainFlow-VLA trajectory is retained as the fallback. Fifteen route-lattice trajectories and their fifteen GETok projections extend its proposal support. A low-capacity rule selected only on legal Trainval logs accepts a GETok projection when its raw/GETok pair agrees on scorer margin and safety components:

consensus learned-score gap >= 0.3
paired ego-progress score    >= 1.0
paired TTC score             >= 3.0

The final overlay changes four of 12,146 navtest trajectories and leaves the other 12,142 predictions from the released ChainFlow-VLA submission intact. The frozen rule is stored in artifacts/g245_frozen_gate.json.

Reproduction

Install the official NAVSIM v1.1 environment and ChainFlow-VLA release first:

The scripts under code/ implement grouped policy selection, policy freezing, posthoc application, immutable fallback overlay, and metadata-only submission finalization. They expect the candidate-score and rank-dump schemas produced by ChainFlow-VLA with 64 base proposals plus 15 raw lattice and 15 GETok proposals.

python code/apply_chainflow_paired_consensus.py --help
python code/overlay_chainflow_submission.py --help
python code/finalize_navsim_submission_metadata.py --help

The uploaded submission.pkl is the exact official candidate. Its trajectory digest and source-submission SHA-256 are recorded in artifacts/g247_official_manifest.json and artifacts/g247_overlay_manifest.json.

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 NAVSIM score is 0.9492706849694933, compared with 0.9491968762246324 for the exact released ChainFlow-VLA submission on the same local cache. The paired delta is +7.38087448606727e-05; its 20,000-repeat bootstrap interval is [-1.9995823431698164e-05, 0.0002457325116051862]. These local values are not presented as an official leaderboard result. The official server score should be used for leaderboard claims.

Authors

Li Ming, Wuhan University (WHU). Contact: Liming751218@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.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support