checkpoints/README.md

TriDrive checkpoints (supplementary)

Only TriDrive's own models and our reproduction of the published Driver-WM pipeline are included. Third-party weights (V-JEPA 2, YOLO11x-pose, RTMW-x-l, MotionBERT release, OpenBADAS) are not redistributed; other baselines (ST-GCN, SiMLPe, MotionBERT fine-tunes, Driver-WM architecture adapted to our tokens) and ablation arms (independent branches, no links, all-to-all, visual-only, sensor-only, road-to-driver) are available on request. Internal run names keep the development code name cwm3d / CockpitWM; they denote the same model as TriDrive in the paper.

Each run directory contains the checkpoint (ckpt_best.pt or ckpt_ep*.pt), the run manifest (run_*_latest.json / config.json: full command line and hyper-parameters), training history and, where the run was evaluated, the test/validation metrics that the paper quotes (the compact forecasters KPF_* and the pose-pretraining checkpoint ship history and manifests only). Absolute paths in manifests and inside the pickled args / cfg dictionaries were rewritten to /home/USER/... (see EXPORT_NOTE*.json); the host name was replaced by REDACTED. Licence: CC BY 4.0 (LICENSE); checksums in MANIFEST.txt.

aide/ — AIDE keypoint forecasting (Table 1, Table 2, Appendix C)

Directory	Paper row	Seeds
cwm3d_kp_s{42,123,456,789,1011}_v2	Ours (5.77M), no pretraining: 48.60 ± .12 All-MPJPE	5
cwm3d_kp_batonpre_trunkp123_s{42,123,456,789,1011}_v2	Ours + BATON pose pretraining: 48.05 ± .06	5
cwm3d_kp_batonpre_pretrain_s123	The BATON pose-pretraining checkpoint whose trunk is transferred (epoch 4 of 6)	1
cwm3d_kp_abs_s{42,123,456}_v2	Absolute-coordinate ablation of the same model (3.87 px worse)	3
cwm3d_vis_sema_vge_s{42,123,456}_v2	View-gate semantic variant (73.5M): ckpt_best_f1.pt = macro-F1-selected row of Table 2; ckpt_best.pt = geometry-selected row of Appendix C	3
driverwm_repro_s{42,123,456}_v2	Our reproduction of the published Driver-WM pipeline (123.7M): 130.86 ± 3.11 All-MPJPE, 194.34 HM h=5	3

baton/ — BATON joint forecasting (Table 3, Appendix B)

Directory	Role
I1_all_minus_road2drv_s{0,1,2}_final_ecan	Directed joint model on ECAN inputs without forecast features: the "Directed (default)" / "ECAN, directed (default)" columns of Table 3 A–C
I1_all_minus_road2drv_s{0,1,2}_final_ecan_kpf	Default registered joint model with the frozen forecast features injected: the "Default" column of Table 3 D
U0_kin_can_noK_s{0,1,2}_kcv2_clip8_ln	Pretrained driver branch (GRU over 18-d kinematics + CAN)
V0_can4s_s{0,1,2}_stats2	Pretrained vehicle branch (causal TCN over 4 s of CAN)
E0_pooled_P_can_s{0,1,2}_v2	Pretrained road branch on ECAN inputs (causal Transformer over 16 V-JEPA 2 tokens + side MLP)
KPF_mb_full_s{0,1,2}	Compact keypoint forecaster (0.85M) that produces the 10-d forecast features $f_t$; s0 is the deployed one. Architecture: the MotionBERT DSTformer re-instantiated and trained from scratch (no released weights; Apache-2.0 attribution in ../THIRD_PARTY_NOTICES.md)
KPF_cwm_full_s{0,1,2}	Alternative compact forecaster architecture evaluated in Appendix B
warning_head/	Training logs, cross-validation tables and selected configurations of the current-state warning probe (histogram gradient boosting over 77 features), including the sub-runs fitted on ablation trunks; the fitted probe itself is deployed/readout/readout_cur_dwfc_s0.pkl
warning_trunk_heads/	Latent-state warning probe experiment of Appendix D

deployed/ — the on-vehicle configuration (Section 3.4, Appendix G)

File	Role
onnx/joint_final_ecan_kpf_s0.onnx	Joint model as executed on the chestnut eGPU (2.35 Hz)
onnx/kpf_mb_full_s0.onnx	Compact keypoint forecaster as executed on the vehicle
onnx/student_v5/	Distilled ViT-S/16 road student (v5, deployed): pooled road feature + risk logit. Derivative of V-JEPA 2 and OpenBADAS; check both licences before redistribution
aux/	Feature standardization statistics and the PCA projections (256-d road tokens, 256-d visual side vector) fitted on the training split
readout/	The deployed HGB probe (readout_cur_dwfc_s0.pkl), its 77 feature names, fill constants, and the validation-fitted thresholds