README.md

TriDrive — code, checkpoints, benchmark tables, derived data, and user-study results

TriDrive is a unified framework that jointly models the driver, the vehicle, and the road for two applications: multi-step forecasting of driver, vehicle, and road-demand states, and a real-time current-state warning probe for driver monitoring. Three modality-specific predictors (anchored driver kinematics from a frozen pose estimator, CAN-bus vehicle dynamics, and frozen V-JEPA 2 road latents with structured road margins) are connected by directed residual connections into the vehicle state and decode 25 future steps (0.2 s) of transition hazards, continuous vehicle dynamics, and future road latents directly from the encoded history. The framework is evaluated on the public AIDE benchmark, on the BATON naturalistic driving corpus, and on a comma four with a chestnut eGPU in a paired on-road user study with 14 drivers.

This bucket accompanies an anonymous submission under double-blind review. Internal identifiers keep the development code name CockpitWM / cwm3d; they denote the same model as TriDrive in the paper.

Contents

Path	What it is	Licence
code/	Training / evaluation code snapshot (AIDE forecaster, joint model, warning probe, baselines, table generation). Layout in code/README.md	MIT (LICENSE)
checkpoints/	TriDrive checkpoints and our Driver-WM reproduction: aide/ (keypoint forecaster, 5+5 seeds, pretraining trunk, ablations, semantic variant, Driver-WM reproduction), baton/ (joint model, branch predictors, compact keypoint forecaster, warning probe), deployed/ (on-vehicle ONNX graphs, PCA/stats, probe). Mapping to paper rows in checkpoints/README.md; checksums in checkpoints/MANIFEST.txt	CC BY 4.0 (checkpoints/LICENSE)
benchmark/labels/	Warning-probe label tables: 979 training anchors (562 manual / 417 engaged), the 512-row gold set and the 509-row headline gold set	CC BY-NC 4.0
benchmark/human_eval_299/	The 299-clip human evaluation: clip manifest, three-stage labels, frozen thresholds, pre-registration (byte-identical, hash included) and result tables	CC BY-NC 4.0
benchmark/aide_protocol/	AIDE split files, protocol and metric code, protocol verification record	MIT / AIDE terms for the data itself
benchmark/tables/	Every result table behind the paper: release tables (AIDE, BATON, driver-monitoring baselines), the generated LaTeX tables, run ledgers, Driver-WM audit and quantile sweeps, EMA confound check	CC BY 4.0
data/	Pseudonymised derived data used by the joint model: anchor grids with CAN histories and targets, road-token bank and anchor index, road PCA packs, ECAN margins and road-dynamics targets, demand / pressure / response targets, PCA matrices. Details in data/README.md	CC BY-NC 4.0 (data/LICENSE-DATA, inherited from BATON)
questionnaire/	The fielded questionnaire (blank), the 14 completed forms anonymised as driver01–driver14 (= P01–P14 in the paper), per-item answers, per-participant construct scores, endpoint statistics, analysis scripts and the figure sources	CC BY 4.0 (questionnaire/LICENSE)
THIRD_PARTY_NOTICES.md	Licences of the components TriDrive builds on	—

Identifier scheme

BATON devices and routes appear only as pseudonyms: driver_N for a device, driver_N__route_M__route_a-b for a route segment group (route_a-b = the segment range of the original recording); a recording id that could not be resolved to a route group appears as a date-free rid_<10 hex>. The pseudonyms are consistent across every file in this bucket and with the public BATON release. Raw device identifiers, recording timestamps, file-system paths, hosts and annotator identities were removed; where a value had to be kept for joins it is the pseudonym.

What is not included

• Driver-side cabin-derived features (kinematics, face/head-pose channels, keypoint residual banks). They are derivatives of in-cabin video of BATON drivers and are withheld pending consent review; the code in code/data/ rebuilds them from the BATON corpus.
• Raw recordings. No BATON or AIDE video, CAN logs, pose data or frames; no on-road recordings of the user study.
• Hold-out device packs (60 additional BATON devices) and per-sample prediction dumps of the joint model (≈ 21 GB); the metrics computed from them are in benchmark/tables/. Available on request after the review period.
• AIDE-derived caches and per-clip predictions (AIDE is obtained from its authors; benchmark/aide_protocol/ rebuilds them).
• Third-party weights (V-JEPA 2, OpenBADAS, YOLO11x-pose, RTMW, MotionBERT release weights); see THIRD_PARTY_NOTICES.md.
• Baseline and ablation checkpoints other than the Driver-WM reproduction (ST-GCN, SiMLPe, MotionBERT fine-tunes, independent / no-link / all-to-all / visual-only / sensor-only arms): available on request.

Headline numbers (from the paper; every table is under benchmark/tables/)

Result	Where
AIDE keypoint forecasting: 48.60 ± 0.12 All-MPJPE without pretraining, 48.05 ± 0.06 with BATON pose pretraining (5 seeds each)	Table 1, Appendix C; checkpoints/aide/
BATON joint forecasting: directed residual connections and road margins vs. independent branches, with paired driver-cluster bootstrap intervals	Table 3, Appendix B; checkpoints/baton/, benchmark/tables/
Warning probe: AUROC 0.725 (manual) vs. 0.563 for the openpilot-based baseline on the human-labelled gold set	Section 4.4, Appendix D; benchmark/labels/
Human evaluation of 299 clips, pre-registered	Appendix E; benchmark/human_eval_299/
On-vehicle: pose 10 Hz, keypoint forecaster and joint model 2.35–5 Hz, end-to-end p95 177 ms on a comma four + RX 9060 eGPU	Section 3.4, Appendix G; checkpoints/deployed/
Paired on-road user study, 14 drivers: appropriateness +1.79 [0.86, 2.75] (n = 14), timeliness +2.67 [1.33, 4.08] (n = 12), annoyance +1.17 [0.33, 2.08] (n = 12); preference 7 : 4 : 3 (TriDrive : stock : no difference)	Section 4.4, Appendix H; questionnaire/

Reproducing the tables

• AIDE (Table 1, Appendix C): code/aide/train/train_kp.py trains one arm per seed; code/aide/eval/ computes All-MPJPE, PCK@0.05 and the high-motion quantile sweeps. Checkpoints under checkpoints/aide/ are the exact runs quoted in the paper.
• BATON (Table 3, Appendix B): code/train/ trains the branch predictors and the joint model on the packs in data/ (driver-side features must be rebuilt first, see above); code/eval/ produces the onset PR-AUC tables with paired driver-cluster bootstrap intervals from checkpoints/baton/.
• Warning probe (Section 4.4, Appendix D): code/eval/ and code/baselines_dms/ evaluate the probe (checkpoints/baton/warning_head/, checkpoints/deployed/readout/) on benchmark/labels/gold_509.csv.
• Human evaluation (Appendix E): benchmark/human_eval_299/ with code/eval/newtest_analysis.py.
• User study (Appendix H): questionnaire/scripts/analysis_questionnaire_v1.py on questionnaire/results/questionnaire_wide.csv.

Environment

Python 3.12, PyTorch, ONNX Runtime, scikit-learn, huggingface_hub for the frozen encoders. Paths in the scripts use the placeholder prefix /home/USER/... (the code resolves its own tree relative to __file__; AIDE_ROOT is read from the environment); replace the placeholders that point at data locations before running. Hosts, IP addresses and the remote data server referenced by the original training-queue scripts were replaced by placeholders.

Licence

Code: MIT (LICENSE). Model weights and questionnaire materials: CC BY 4.0. Derived data and labels: CC BY-NC 4.0, inherited from the BATON corpus. Third-party components: see THIRD_PARTY_NOTICES.md. Attribution: "TriDrive authors (anonymous for double-blind review)"; the camera-ready version will add the citation.