File size: 36,967 Bytes
6620fae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | # NeDM traversing release
Data and models of the traversing study (learned route-risk planning for off-road vehicles in Project
Chrono). The full description is in Part B of the dataset card at the repository root; every file with
its size and SHA256 is in `release_manifest.json`. Restore items into a NeDM checkout with
`python traversing/scripts/release/download_traversing_data.py --list --all` and `--items NAME ...`.
Source commit: `901d6c9423a16c0fafc3d60056065415d5a725f2`.
| item | bundle | milestone | packing | files | bytes | restores to | description |
|---|---|---|---|---|---|---|---|
| `arena_gator_collection_task_files` | raw | m4a, m4b | files | 4 | 102,531,437 | `artifacts/traverse/arena_gator_20260925/e3/tasks` | Task files passed to the collection builds of the arena and vehicle studies: the Gator rigid pool (rigid_v2), the Gator soil builds (soil_v3 for tiers 0-12, soil_v2 for tiers 0-6, which also defines the HMMWV g203/g228 soil tiers) and the HMMWV g203/g228 rigid pools (rigid_hmmwv_v1). |
| `arena_heightmaps` | assets | m1, m2, m3, m4a, m4b | files | 37 | 4,792,301 | `assets/traverse` | Heightmap and metadata of every arena driven in a milestone: f104 and 17 generated arenas, plus the arena family description. |
| `crm_soil_config_and_vehicle_variants` | assets | m2, m3, m4a, m4b | files | 14 | 42,010 | `.` | The single deformable-soil setting (crm_main.json), the M113 smoke-test soil setting (crm_m113.json) and the Polaris and M113 vehicle definition variants. |
| `f104_eval_suite_800` | evaluation | m2, m4a, m4b | tar | 2 | 5,680,163 | `artifacts/traverse/generalist_20260921/A_adapt/suite` | The frozen 800-pair f104 start/goal suite: suite definition, lock file, 800 case files, the designed route of each group and the locked standing-start task lists. |
| `f104_reference_routes_and_cases` | evaluation | m3, m4b | tar | 2 | 1,410,474 | `artifacts/traverse/fdm_f104_50h_20260909` | f104 start/goal case files and reference routes that released drive tasks point at: the 423 routes and 55 cases the tracker drives follow, and the 147 routes (88 designed, 59 on-policy) and 27 cases of the vehicle smoke test: the 144 sample-A routes (86 designed, 58 planner proposals) plus the 3 routes of the bit-identity check. |
| `f104_terrain_grid` | assets | m3 | files | 1 | 3,267,723 | `artifacts/traverse/crm_f104_v1/grids/arena_f104_50h_v1` | Metric terrain grid of f104 used for the terrain crops of the vehicle-dynamics model and the tracker training environment. |
| `gator_collection_qa_records` | raw | m4b | files | 8 | 9,146,964 | `artifacts/traverse/arena_gator_20260925/e5` | Recorded QA of the Gator collections and the validated-id lists passed to the subset builder: soil all tiers, soil tiers 0-6 and rigid. |
| `gator_f104_rigid_collection_runs` | raw | m4b | tar | 6 | 1,978,884,155 | `artifacts/traverse/arena_gator_20260925/rigid_v1/runs` | The Gator driving the 24,000-route rigid f104 pool (1,200 groups x 20: 14,400 designed routes and 9,600 planner-style routes), 165.7 simulated hours. Chain: ci_f104_gator_rigid.npz (93,551 rows), then subset G_f104_gator_rigid.npz, then the G_rigid ensemble (e5/deploy/G). |
| `gator_f104_soil_collection_runs` | raw | m4b | tar | 6 | 2,316,775,153 | `artifacts/traverse/arena_gator_20260925/soil_v1/runs` | The Gator driving the 15,235 HMMWV f104 soil route ids (tiers 0-12: 9,168 designed routes and 6,067 planner proposals), 141.1 simulated hours. These are the source of ci_f104_gator_crm.npz (tiers 0-12, 55,826 rows), which feeds the G_full_soil ensemble. They are also the source of the stage-1 file (tiers 0-6, 8,399 runs, 30,827 rows), which feeds the G_soil ensemble. |
| `gator_rigid_f104_drive_folders` | evaluation | m4b | tar | 3 | 68,684,825 | `artifacts/traverse/arena_gator_20260925/e6/runs_rigid` | Drive folders of the Gator rigid f104 read-out in the README bar table: the Gator-trained planner (800/800) and the straight route at 6 m/s. |
| `gator_rigid_subset_all` | processed | m4b | files | 1 | 1,697,443,990 | `artifacts/traverse/arena_gator_20260925/e4/subsets` | Gator rigid on f104, all 24,000 route ids. 93,551 rows, 84,922 fitted. Trained G, the Gator rigid planner cited in the README bar table (800/800). |
| `gator_soil_subset_all` | processed | m4b | files | 1 | 1,005,085,677 | `artifacts/traverse/arena_gator_20260925/e4/soil_bf/subsets` | Gator soil on f104, all 15,235 validated ids (tiers 0-12). 55,826 rows, 50,822 fitted. Trained G_full_soil (columns gator_own_model_sampling, gator_own_model_sampling_grad). |
| `gator_soil_subset_tiers0to6` | processed | m4b | files | 1 | 557,066,585 | `artifacts/traverse/arena_gator_20260925/e4/soil_s1/subsets` | Gator soil on f104, tiers 0-6 (8,399 drives). 30,827 rows, 28,057 fitted. Trained G_soil (column gator_own_model_tiers0to6_sampling). |
| `hmmwv_rigid_f104_branch_anchor_records` | raw | m2 | tar | 3 | 518,852 | `artifacts/traverse/generalist_20260921/rigid_a4/runs` | Per-anchor records of the HMMWV rigid f104 branch drives (<episode>__a4, 800 anchor states): the automatic branch choice (branch_auto.json, 789) or the skip record (skipped.json, 11) and the completion marker. The drives themselves are hmmwv_rigid_f104_branch_continuations. |
| `hmmwv_rigid_f104_branch_continuations` | raw | m2 | tar | 4 | 250,801,904 | `artifacts/traverse/generalist_20260921/rigid_a4/runs` | HMMWV rigid drives that branch from a moving state on f104: 2,367 continuation drives (<anchor>__a4__c<j>; 2,358 produce rows, 9 were skipped) from 800 anchor states. The per-anchor records are hmmwv_rigid_f104_branch_anchor_records. |
| `hmmwv_rigid_f104_early_waves` | raw | m1, m2 | tar | 4 | 84,885,327 | `artifacts/traverse/fdm_f104_50h_20260909` | Seven small early HMMWV rigid f104 waves whose drives are in the navigation model's 36,199-route f104 label file: prospective_v1 309, fixed_speed_v1 292, speed_c4 206, wide_v1 146, speed_c6 90, band_v1 73, showcase_v1 14. The last six make up the 821 rows the label file tags as MPPI arms. |
| `hmmwv_rigid_f104_first_overnight_collection` | raw | m1, m2 | tar | 4 | 681,312,399 | `artifacts/traverse/fdm_f104_50h_20260909/production_v2/runs` | Earlier HMMWV rigid f104 production collection (1,500 groups, the first overnight wave). It is not part of the 24,000-route pool. It supplies 11,412 of the navigation model's 36,199 f104 routes and part of the rigid N2 specialist's training rows. |
| `hmmwv_rigid_f104_pool_designed` | raw | m1, m2, m3, m4a | tar | 4 | 852,972,034 | `artifacts/traverse/fdm_f104_50h_20260909/production_v3/runs` | HMMWV rigid-ground pool on f104, designed routes: 1,200 start/goal groups x 12 routes. This is the first half of the 24,000-route rigid pool. Every later rigid f104 model and the tracker cache were built from this pool, and the Gator and Polaris re-drove its route ids. |
| `hmmwv_rigid_f104_pool_onpolicy` | raw | m1, m2, m3, m4a | tar | 5 | 1,088,263,644 | `artifacts/traverse/fdm_f104_50h_20260909/production_v4/runs` | HMMWV rigid-ground pool on f104, planner-style (on-policy) routes: 1,200 groups x 8 routes. This is the second half of the 24,000-route rigid pool. |
| `hmmwv_rigid_g203_pool` | raw | m4a | tar | 6 | 2,048,751,721 | `artifacts/traverse/arena_gator_20260925/rigid_v1/runs` | HMMWV rigid 24,000-route pool on the extra training arena g203 (1,200 groups x 20 routes; 14,400 route_ ids + 9,600 op_ ids). ag_build_ds.py selected all 24,000 (92,225 rows). |
| `hmmwv_rigid_g228_pool` | raw | m4a | tar | 6 | 1,986,566,300 | `artifacts/traverse/arena_gator_20260925/rigid_v1/runs` | HMMWV rigid 24,000-route pool on the extra training arena g228 (93,113 rows). |
| `hmmwv_rigid_multi_arena_test_drives` | raw | m1 | tar | 4 | 250,335,033 | `artifacts/traverse/fdm_f104_50h_20260909/gen_v1/test/runs` | HMMWV rigid test drives on f104, g203, g217 and g228 from the multi-arena study (planner, rule-based, fixed-speed and straight-route arms; 1,103 / 1,098 / 1,132 / 1,111). The navigation trainer (sensor_train_v2.py) fits every row from a training arena whatever its source, so these are training rows. |
| `hmmwv_rigid_sensor_study_test_drives` | raw | m1 | tar | 4 | 408,101,749 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v1` | HMMWV rigid drives from the sensor study's two test campaigns (sensor_v1/test 3,614 and test2 4,523) on the training arenas f104, g203, g217 and g228. They enter the navigation training rows through ds_v2_gen_<arena>.npz. |
| `hmmwv_rigid_sibling_arenas_designed_routes` | raw | m1 | tar | 4 | 329,701,570 | `artifacts/traverse/fdm_f104_50h_20260909/gen_v1/data/runs` | HMMWV rigid designed-route drives on the navigation model's three sibling training arenas g203, g217 and g228 (1,800 each, 5,400 total). The g216 and g231 drives in the same folder are held-out evaluation only and are excluded. |
| `hmmwv_rigid_tracker_held_control_drives` | raw | m3 | tar | 4 | 100,479,176 | `artifacts/traverse/generalist_20260921/rigid_b3/runs` | HMMWV rigid drives with perturbed controls on f104 training routes: controls held for a true 50 ms with brake taps added (1,500 drives, <route>__b3). |
| `hmmwv_rigid_tracker_policy_harvest` | raw | m3 | tar | 4 | 68,924,032 | `artifacts/traverse/generalist_20260921/rigid_harvest/runs` | HMMWV rigid drives by the round-1 learned tracker on 1,000 training routes (<route>__harvest). They were added to the dynamics cache. |
| `hmmwv_soil_f104_branch_continuations` | raw | m2 | tar | 4 | 246,449,442 | `artifacts/traverse/crm_f104_v1/generalist/a4/pass2_out/runs` | HMMWV soil drives that branch from a moving state on f104: 752 anchor states x 3 = 2,256 continuation drives (<episode>__c<j>). |
| `hmmwv_soil_f104_collection` | raw | m2, m3, m4a, m4b | tar | 5 | 1,415,773,693 | `artifacts/traverse/crm_f104_v1/collect_v1/runs` | HMMWV soil collection on f104 (Chrono's particle-based deformable soil, 0.24 m deep over a rigid floor): 15,235 validated route ids in 1,200 groups, tiers 0-12 (9,168 designed + 6,067 planner-style). The Gator and Polaris later re-drove the same ids. |
| `hmmwv_soil_g203_tiers0to6` | raw | m4a | tar | 4 | 384,588,689 | `artifacts/traverse/arena_gator_20260925/soil_v1/runs` | HMMWV soil drives on g203, tiers 0-6 (the first 7 routes of each of 606 start/goal groups): 4,242 driven, of which 4,240 pass QA (2 rejected as crm_qa:unstalled_break). |
| `hmmwv_soil_g228_tiers0to6` | raw | m4a | tar | 4 | 376,774,712 | `artifacts/traverse/arena_gator_20260925/soil_v1/runs` | HMMWV soil drives on g228, tiers 0-6: 4,242 driven, all validated. |
| `hmmwv_soil_subset_gator_matched_all` | processed | m4b | files | 1 | 1,055,259,195 | `artifacts/traverse/arena_gator_20260925/e4/soil_bf/subsets` | HMMWV soil on f104, exactly the route ids the Gator validated (all tiers). 58,268 rows, 52,923 fitted. Trained H_full_soil (columns gator_hmmwv_model_sampling, hmmwv_own_model_sampling, hmmwv_own_model_sampling_grad). |
| `hmmwv_soil_tracker_held_control_drives` | raw | m3 | tar | 4 | 129,549,038 | `artifacts/traverse/crm_f104_v1/generalist/mixed_v1/out/runs` | HMMWV soil drives with the same held-control perturbations on f104 training routes (1,500 drives, <route>__b3). |
| `m1_nav_direct_depth_risk_ensemble` | models | m1 | files | 4 | 3,128,255 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2/matched` | Navigation risk ensemble: 3 seeds (s0-s2) of the direct-depth corridor model (channels range_abs, sec1, speed, valid) that scored every candidate route in all 120 navigation missions (all 4 decision-timing arms). |
| `m2_joint_station_history_transformer` | models | m2 | files | 6 | 8,416,069 | `artifacts/traverse/crm_improve_20260922/deploy_v1` | Joint station+history transformer trained on the same rows (deploy_a3_haux_txjoint, 5 seeds s0-s4) behind transformer_1s, transformer_0p5s and transformer_0p5s_grad (soil). |
| `m2_oracle_world_label_model` | models | m2 | files | 6 | 5,322,332 | `artifacts/traverse/generalist_20260921/A_adapt/train/deploy_v1` | Shared model given the true world label (T, 5 seeds s0-s4) behind oracle_tag_standing, oracle_tag_3s, oracle_tag_1s, oracle_tag_0p5s. |
| `m2_pooled_no_label_no_history_model` | models | m2 | files | 6 | 5,321,276 | `artifacts/traverse/generalist_20260921/A_adapt/train/deploy_v1` | Shared model with neither history nor label (P, 5 seeds s0-s4) behind pooled_3s, pooled_1s, pooled_0p5s. |
| `m2_rigid_only_planner_standing_start` | models | m2 | files | 6 | 5,195,180 | `artifacts/traverse/fdm_f104_50h_20260909/night2_v1/final` | Earlier deployed rigid-only planner (night-2 N2, 5 seeds s0-s4) behind column specialist_rigid_standing (study arm Srigid, sampling arm B) in both m2 tables. |
| `m2_rigid_specialist_same_rows` | models | m2 | files | 6 | 5,253,018 | `artifacts/traverse/generalist_20260921/A_adapt/train/deploy_v1` | Single-world rigid model trained on the same re-anchored rows (Sp_rigid, 5 seeds s0-s4) behind specialist_rigid_3s. |
| `m2_shared_history_early_rows_final_model` | models | m2 | files | 6 | 6,203,497 | `artifacts/traverse/crm_improve_20260922/deploy_v1` | Model of the FINAL label-free configuration: the history design retrained with extra early-decision and moving-branch rows (deploy_a1_haux_gru, 5 seeds s0-s4). Behind shared_hist_early_rows_1s, shared_hist_early_rows_0p5s (soil and rigid) and shared_hist_early_rows_0p5s_grad (the 780/800 soil and 800/800 rigid headline). |
| `m2_shared_history_model` | models | m2 | files | 6 | 5,452,376 | `artifacts/traverse/generalist_20260921/A_adapt/train/deploy_v1` | Shared rigid/soil history model H (CNN-GRU + 2 s history encoder, 5 seeds s0-s4) behind shared_hist_standing, shared_hist_3s, shared_masked_3s (history masked at decision), shared_hist_1s and shared_hist_0p5s. |
| `m2_soil_only_planner_standing_start` | models | m2 | files | 6 | 5,216,962 | `artifacts/traverse/crm_f104_v1/train_v1/deploy` | Earlier deployed soil-only planner (CRM_N2, 5 seeds s0-s4) behind column specialist_soil_standing (study arm Scrm, sampling arm B) in both m2 tables. |
| `m2_soil_specialist_same_rows` | models | m2 | files | 6 | 5,254,231 | `artifacts/traverse/generalist_20260921/A_adapt/train/deploy_v1` | Single-world soil model trained on the same re-anchored rows (Sp_crm, 5 seeds s0-s4) behind specialist_soil_3s and specialist_soil_1s. |
| `m3_tracker_learned_dynamics_model` | models | m3 | files | 2 | 19,658,720 | `artifacts/traverse/generalist_20260921/B_tracker/nrd_tag_v3` | Tag-conditioned NRD vehicle-dynamics model (causal transformer, 17-D state, 50 ms steps, 8x8 terrain crop) inside which the round-2 PPO policy was trained; not used when driving. |
| `m3_tracker_round2_numpy_actor` | models | m3 | files | 2 | 2,009,045 | `artifacts/traverse/generalist_20260921/B_tracker/ppo_v2` | Deployed round-2 learned tracker: the PPO actor exported to NumPy (158-D observation, 3 actions), run without torch or the NRD model; behind column nrd_policy_v2 of m3_tracker_routes.csv. |
| `m3_tracker_round2_ppo_checkpoint` | models | m3 | files | 3 | 5,930,492 | `artifacts/traverse/generalist_20260921/B_tracker/ppo_v2` | Final PPO torch checkpoint (iteration 999) of the round-2 tracker, the source of the NumPy actor. Not needed to re-drive (the actor is enough) and not a training input, but it is the only artefact with which the actor export and parity check can be repeated. |
| `m4a_rigid_f104_only_ens1` | models | m4a | files | 8 | 7,792,049 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M1a` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M1a_rigid_deploy, behind columns f104_only_ens1_fixed2mps and _speedfree (rigid, M1a). The same bytes are aliased as H, the HMMWV-trained rigid model of the Gator study. |
| `m4a_rigid_f104_only_ens2` | models | m4a | files | 8 | 7,794,713 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M1b` | CNN-GRU route-risk ensemble, 5 seeds (s5-s9), M1b_rigid_deploy, behind columns f104_only_ens2_fixed2mps and _speedfree (rigid, M1b). |
| `m4a_rigid_three_arenas_all_data` | models | m4a | files | 8 | 12,462,649 | `artifacts/traverse/arena_gator_20260925/e5/deploy/A3` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), A3_rigid_deploy, behind columns three_arenas_all_data_fixed2mps and _speedfree (rigid, A3). |
| `m4a_rigid_three_arenas_same_total_ens1` | models | m4a | files | 8 | 8,028,285 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M3a` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M3a_rigid_deploy, behind columns three_arenas_same_total_ens1_fixed2mps and _speedfree (rigid, M3a). |
| `m4a_rigid_three_arenas_same_total_ens2` | models | m4a | files | 8 | 8,030,122 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M3b` | CNN-GRU route-risk ensemble, 5 seeds (s5-s9), M3b_rigid_deploy, behind columns three_arenas_same_total_ens2_fixed2mps and _speedfree (rigid, M3b). |
| `m4a_rigid_two_arenas_same_total` | models | m4a | files | 8 | 7,896,875 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M2` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M2_rigid_deploy, behind columns two_arenas_same_total_fixed2mps and _speedfree (rigid, M2). |
| `m4a_soil_f104_only_ens1` | models | m4a | files | 8 | 6,269,631 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M1a_soil` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M1a_soil_deploy, behind column f104_only_ens1 (soil table, study M1a). The same bytes are aliased as H_soil, the Gator study's stage-1 HMMWV soil model. |
| `m4a_soil_f104_only_ens2` | models | m4a | files | 8 | 6,269,028 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M1b_soil` | CNN-GRU route-risk ensemble, 5 seeds (s5-s9), M1b_soil_deploy, behind column f104_only_ens2 (soil table, study M1b). |
| `m4a_soil_three_arenas_all_data` | models | m4a | files | 8 | 7,089,690 | `artifacts/traverse/arena_gator_20260925/e5/deploy/A3_soil` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), A3_soil_deploy, behind column three_arenas_all_data (soil, A3). |
| `m4a_soil_three_arenas_same_total_ens1` | models | m4a | files | 8 | 6,312,040 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M3a_soil` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M3a_soil_deploy, behind column three_arenas_same_total_ens1 (soil, M3a). |
| `m4a_soil_three_arenas_same_total_ens2` | models | m4a | files | 8 | 6,311,730 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M3b_soil` | CNN-GRU route-risk ensemble, 5 seeds (s5-s9), M3b_soil_deploy, behind column three_arenas_same_total_ens2 (soil, M3b). |
| `m4a_soil_two_arenas_same_total` | models | m4a | files | 8 | 6,279,922 | `artifacts/traverse/arena_gator_20260925/e5/deploy/M2_soil` | CNN-GRU route-risk ensemble, 5 seeds (s0-s4), M2_soil_deploy, behind column two_arenas_same_total (soil, M2). |
| `m4b_gator_rigid_own_model` | models | m4b | files | 8 | 7,630,859 | `artifacts/traverse/arena_gator_20260925/e5/deploy/G` | Gator-trained rigid ensemble on the Gator's 24,000 f104 rigid drives (CNN-GRU, 5 seeds s0-s4, G_rigid_deploy). Behind the README bar-table Gator rigid row (800/800 on f104; arms G_free at free speed and the G_fixed2 addendum); not a compact-table column. |
| `m4b_gator_soil_own_model_full_data` | models | m4b | files | 8 | 6,842,395 | `artifacts/traverse/arena_gator_20260925/e5/deploy/G_full_soil` | Gator-trained soil ensemble on all 15,235 Gator f104 soil drives (CNN-GRU, 5 seeds s0-s4, G_full_soil_deploy), behind gator_own_model_sampling (Gfull_free_gator) and gator_own_model_sampling_grad (Gfull_grad_gator). |
| `m4b_gator_soil_own_model_tiers0to6` | models | m4b | files | 8 | 6,217,551 | `artifacts/traverse/arena_gator_20260925/e5/deploy/G_soil` | Gator-trained soil ensemble on only the first 7 routes of each group (8,399 drives; CNN-GRU, 5 seeds s0-s4, G_soil_deploy), behind gator_own_model_tiers0to6_sampling (G_free_gator). |
| `m4b_hmmwv_soil_model_full_data` | models | m4b | files | 8 | 6,933,125 | `artifacts/traverse/arena_gator_20260925/e5/deploy/H_full_soil` | HMMWV-trained soil ensemble on the full HMMWV f104 soil file, the same ids as the Gator (CNN-GRU, 5 seeds s0-s4, H_full_soil_deploy). Behind gator_hmmwv_model_sampling (Hfull_free_gator), hmmwv_own_model_sampling (Hfull_free) and hmmwv_own_model_sampling_grad (Hfull_grad_hmmwv). |
| `m4b_polaris_soil_own_model_full_data` | models | m4b | files | 7 | 5,436,589 | `artifacts/traverse/offroad_vehicles_20260927/e5/deploy/polaris_full_soil` | Final Polaris soil ensemble on all 15,229 valid Polaris f104 soil drives (CNN-GRU, 5 seeds s0-s4, polaris_full_soil_deploy). Behind polaris_own_model_sampling_grad and polaris_own_model_sampling on f104 and on the 8 unseen arenas; polaris_corrected_driveline_grad_routes re-drives the polaris_grad routes. |
| `nav_corrected_mission_outcomes` | evaluation | m1 | files | 121 | 377,177 | `artifacts/traverse/fdm_f104_50h_20260909/nav_v1/local_luffy` | Corrected navigation read-out after the rescue-route fix: summary.json and the 120 per-mission outcome files (30 missions x 4 decision-timing arms) behind m1_navigation_missions.csv. |
| `nav_corrected_run_folders` | evaluation | m1 | tar | 3 | 33,827,079 | `artifacts/traverse/fdm_f104_50h_20260909/nav_v1/local_luffy` | Per-mission decisions, candidate routes and trajectory of the 120 corrected navigation drives. |
| `nav_depth_corridors_arena_g203` | processed | m1 | files | 1 | 365,853,046 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors on sibling training arena g203. 4,925 rows. |
| `nav_depth_corridors_arena_g216_heldout` | processed | m1 | files | 1 | 368,155,850 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors on held-out arena g216 (read by the trainer for its held-out route-choice score, not fitted). 4,956 rows. |
| `nav_depth_corridors_arena_g217` | processed | m1 | files | 1 | 372,018,618 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors on sibling training arena g217. 5,008 rows. |
| `nav_depth_corridors_arena_g228` | processed | m1 | files | 1 | 366,744,454 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors on sibling training arena g228. 4,937 rows. |
| `nav_depth_corridors_arena_g231_heldout` | processed | m1 | files | 1 | 363,178,822 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors on held-out arena g231 (read by the trainer for its held-out score, not fitted). 4,889 rows. |
| `nav_depth_corridors_f104_extra_drives` | processed | m1 | files | 1 | 231,176,534 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Navigation corridors from f104 gen_v1 / sensor_v1 test drives (fitted; f104 is a training arena). 3,111 rows. |
| `nav_depth_corridors_f104_training_routes` | processed | m1 | files | 1 | 2,675,496,171 | `artifacts/traverse/fdm_f104_50h_20260909/sensor_v2` | Direct-depth route-corridor tensors for the navigation risk ensemble (matched_Dabs, 3 seeds): all 36,199 f104 rigid training routes (labels reused from the merged station set via R/sensor_v1/labels_station_ds_all.npz, corridors re-sampled from the back-projected depth grid). 36,199 rows. |
| `nav_missions_and_task_list` | evaluation | m1 | files | 31 | 67,059 | `artifacts/traverse/fdm_f104_50h_20260909/nav_v1` | The 30 frozen waypoint missions and the 120-row task list (mission x decision-timing arm). |
| `planner_map_f104` | evaluation | m2, m4a, m4b | files | 2 | 4,099,355 | `artifacts/traverse/crm_f104_v1/maps/arena_f104_50h_v1` | Whole-arena overhead depth map of f104 on which every f104 planner scores its candidate routes. |
| `planner_maps_new_arenas` | evaluation | m4a, m4b | files | 23 | 45,171,686 | `artifacts/traverse/arena_gator_20260925/maps` | Overhead depth maps of the 11 other planner arenas (g203, g217, g228 and the 8 unseen test arenas) and the map checks. |
| `polaris_collection_build_record` | raw | m4b | files | 2 | 551,823 | `artifacts/traverse/offroad_vehicles_20260927/e4` | The Polaris stage-2 build record (15,235 candidates, 6 rejected, 15,229 selected) and the validated id list passed to the subset builder. |
| `polaris_collection_task_file` | raw | m4b | files | 1 | 15,894,875 | `artifacts/traverse/offroad_vehicles_20260927/tasks` | soil_v5_polaris.json: the task file passed as --tasks-crm to the Polaris stage-2 build. Its tier 0-12 rows are the 15,235 collection ids. |
| `polaris_f104_soil_collection_runs` | raw | m4b | tar | 5 | 1,659,359,903 | `artifacts/traverse/offroad_vehicles_20260927/soil_v1/runs` | The Polaris (stock driveline) driving the same 15,235 f104 soil ids: 15,229 valid, 93.0 simulated hours. Chain: stage-2 ci_f104_polaris_crm.npz (tiers 0-12, 57,228 rows), then polaris_full_f104_soil.npz, then the polaris_full_soil ensemble. The 144 sample-A rows of the smoke arm 'polaris' are part of this collection. |
| `polaris_soil_subset_all` | processed | m4b | files | 2 | 1,042,216,902 | `artifacts/traverse/offroad_vehicles_20260927/e4/subsets` | Polaris soil on f104, all 15,229 valid ids (tiers 0-12). 57,228 rows, 52,021 fitted. Trained polaris_full_soil (all Polaris planner columns on f104 and the unseen arenas). |
| `polaris_unseen_drive_folders` | evaluation | m4b | tar | 3 | 183,050,061 | `artifacts/traverse/offroad_vehicles_20260927/e6/sync/runs_unseen` | Drive folders behind the 3,000 cells of m4_polaris_unseen_soil.csv (2,618 distinct drives). |
| `rigid_specialist_training_rows` | processed | m2 | files | 1 | 1,117,924,155 | `artifacts/traverse/fdm_f104_50h_20260909/night2_v1` | Training rows of the earlier rigid-only f104 planner N2 (column specialist_rigid_standing, arm Srigid): merged night-1 + night-2 rigid route set, 36,199 routes / 2,700 groups (train 33,840, val 764, test 1,595; 31,851 fitted per N2_meta.json). |
| `shared_model_baseline_reanchored_rows` | processed | m2 | files | 1 | 2,100,906,657 | `artifacts/traverse/generalist_20260921/A_adapt/datasets` | Re-anchored rigid + soil rows with a 2 s history on the f104 route ids: 115,868 rows (105,193 fitted). Training file of the generalist ensembles H, P, T, Sp_crm, Sp_rigid (columns oracle_tag_*, shared_hist_standing/3s/1s/0p5s, pooled_*, shared_masked_3s, specialist_soil_3s/1s, specialist_rigid_3s). |
| `shared_model_decision_states` | evaluation | m2 | tar | 2 | 53,455,546 | `artifacts/traverse` | Inputs of the moving-start decisions: straight approach routes, the approach drives, and the recorded decision poses and 2 s history arrays at 3 s, 1 s and 0.5 s, with the approach task files. |
| `shared_model_drive_folders_rigid` | evaluation | m2 | tar | 3 | 434,695,018 | `artifacts/traverse` | Per-drive folders behind every cell of m2_shared_risk_rigid.csv (9,469 distinct drives for 9,600 cells). |
| `shared_model_drive_folders_soil` | evaluation | m2 | tar | 4 | 902,721,438 | `artifacts/traverse` | Per-drive folders behind every cell of m2_shared_risk_soil.csv (18,027 distinct drives for 18,400 cells). |
| `shared_model_drive_task_lists` | evaluation | m2 | files | 14 | 25,822,103 | `artifacts/traverse` | Task lists that drove the milestone-2 table arms (standing start, 3 s, 1 s, 0.5 s and final), with the two standing-start lock files; those two locks cover unreleased route copies and cannot be checked from the release. |
| `shared_model_final_anchor_k60_rows` | processed | m2 | files | 1 | 541,347,614 | `artifacts/traverse/crm_improve_20260922/datasets` | Final shared-model training file 3 of 3: decision rows at frame 60 (3 s), the k=60 slice of anchor_k40_60_80.npz (14,972 rigid + 14,727 soil). 29,699 rows, 26,936 fitted. |
| `shared_model_final_reanchored_plus_branch_rows` | processed | m2 | files | 1 | 2,182,584,035 | `artifacts/traverse/generalist_20260921/A_adapt/datasets` | Final shared-model training file 1 of 3: mixed_reanchor rows plus 4,614 moving-prefix branch rows of both worlds. 120,482 rows, 109,244 fitted. |
| `shared_model_final_short_anchor_rows` | processed | m2 | files | 1 | 1,584,420,109 | `artifacts/traverse/crm_improve_20260922/datasets` | Final shared-model training file 2 of 3: early decision rows at frames 10/20/30 of every rigid and soil f104 episode. 89,998 rows, 81,634 fitted. |
| `shared_model_results_early_and_final` | evaluation | m2 | files | 4 | 16,659,963 | `artifacts/traverse/crm_improve_20260922` | Per-pair results of the 1 s and 0.5 s early-decision arms and of the final gradient-refined configuration on soil and rigid ground. |
| `shared_model_results_standing_and_3s` | evaluation | m2 | files | 4 | 10,644,015 | `artifacts/traverse/generalist_20260921/A_adapt` | Per-pair results of the standing-start arms and the 3 s-approach arms, soil and rigid. |
| `shared_model_route_picks_3s` | evaluation | m2 | tar | 2 | 45,807,257 | `artifacts/traverse/generalist_20260921/A_adapt` | Locked planner output of the 3 s-approach arms. |
| `shared_model_route_picks_early` | evaluation | m2 | tar | 2 | 52,655,716 | `artifacts/traverse/crm_improve_20260922` | Locked planner output of the 1 s and 0.5 s early-decision arms (sampling search only), including the rigid 0.5 s arm. |
| `shared_model_route_picks_final` | evaluation | m2 | tar | 2 | 38,324,071 | `artifacts/traverse/crm_improve_20260922` | Locked planner output of the final gradient-refined arms after the 0.5 s approach (soil and rigid). |
| `shared_model_route_picks_standing` | evaluation | m2 | tar | 2 | 47,610,089 | `artifacts/traverse/generalist_20260921/A_adapt` | Locked planner output of the standing-start arms: per-pair pick records with candidate scores, route files, task list, run summary and lock file. |
| `soil_specialist_training_rows` | processed | m2 | files | 1 | 470,969,671 | `artifacts/traverse/crm_f104_v1/datasets` | Training rows of the earlier soil-only f104 planner CRM_N2 (column specialist_soil_standing, arm Scrm): one row per validated HMMWV soil route (15,235); 13,821 fitted per CRM_N2_deploy.json. |
| `terrain_rigid_subset_f104_only` | processed | m4a | files | 1 | 1,692,950,845 | `artifacts/traverse/arena_gator_20260925/e4/subsets` | HMMWV rigid, f104 only, 1,089 groups. 93,397 rows, 84,787 fitted. Trained M1a and M1b (columns f104_only_ens1/ens2 _fixed2mps/_speedfree); byte-identical to H_f104_hmmwv_rigid.npz (Gator-study H). |
| `terrain_rigid_subset_three_arenas_all` | processed | m4a | files | 1 | 5,046,634,245 | `artifacts/traverse/arena_gator_20260925/e4/subsets` | HMMWV rigid, all data: f104 1,089 + g203 1,083 + g228 1,063 groups. 278,735 rows, 250,490 fitted. Trained A3 (column three_arenas_all_data). |
| `terrain_rigid_subset_three_arenas_same_total` | processed | m4a | files | 1 | 2,040,832,278 | `artifacts/traverse/arena_gator_20260925/e4/subsets` | HMMWV rigid, f104 + g203 + g228, same total. 112,726 rows, 84,481 fitted. Trained M3a and M3b (columns three_arenas_same_total_ens1/ens2). |
| `terrain_rigid_subset_two_arenas` | processed | m4a | files | 1 | 1,847,970,222 | `artifacts/traverse/arena_gator_20260925/e4/subsets` | HMMWV rigid, f104 + g203, same total. 101,916 rows, 84,310 fitted. Trained M2 (column two_arenas_same_total). |
| `terrain_soil_subset_f104_only` | processed | m4a | files | 1 | 583,333,729 | `artifacts/traverse/arena_gator_20260925/e4/soil_s1/subsets` | HMMWV soil, f104 only, 1,089 training groups, tiers 0-6. 32,152 rows, 29,210 fitted. Trained M1a_soil and M1b_soil (columns f104_only_ens1/ens2, soil) and is byte-identical to H_f104_hmmwv_soil.npz (Gator-study H_soil). |
| `terrain_soil_subset_three_arenas_all` | processed | m4a | files | 1 | 1,167,017,074 | `artifacts/traverse/arena_gator_20260925/e4/soil_s1/subsets` | HMMWV soil, all data: f104 1,089 + g203 559 + g228 520 groups. 64,400 rows, 57,898 fitted. Trained A3_soil (column three_arenas_all_data). |
| `terrain_soil_subset_three_arenas_same_total` | processed | m4a | files | 1 | 644,237,311 | `artifacts/traverse/arena_gator_20260925/e4/soil_s1/subsets` | HMMWV soil, f104 + g203 + g228, 363 groups each. 35,573 rows, 29,071 fitted. Trained M3a_soil and M3b_soil (columns three_arenas_same_total_ens1/ens2). |
| `terrain_soil_subset_two_arenas` | processed | m4a | files | 1 | 602,919,881 | `artifacts/traverse/arena_gator_20260925/e4/soil_s1/subsets` | HMMWV soil, f104 + g203, 545 groups each. 33,227 rows, 29,048 fitted. Trained M2_soil (column two_arenas_same_total). |
| `tracker_drive_folders` | evaluation | m3 | tar | 3 | 140,743,979 | `artifacts/traverse` | Drive folders of the 2,538 tracker table rows (native PID, held PID and learned tracker on 423 routes, both grounds). |
| `tracker_dynamics_training_cache` | processed | m3 | tar | 3 | 1,969,817,340 | `artifacts/traverse/generalist_20260921/B_tracker/cache_v3` | Per-episode 50 ms state/action/pose/power cache of 43,235 f104 episodes (rigid 26,500 = 24,000 production + 1,500 perturbed + 1,000 round-1 policy harvest; soil 16,735 = 15,235 collection + 1,500 perturbed) plus cache_manifest.json and build_report.json. |
| `tracker_group_split_file` | processed | m3 | files | 1 | 169,042,919 | `artifacts/traverse/crm_night2_v1/datasets` | Night-2 soil twin dataset whose group/split arrays define the 1,200-group split (1,089/56/55) used by the cache builder and cross-checked by the NRD trainer. |
| `tracker_results_suite_tasks` | evaluation | m3 | files | 7 | 6,190,906 | `artifacts/traverse/generalist_20260921/B_tracker` | Per-route tracker results (both grounds), the 423-route tracking suite with strata, and the task lists that drove the PID and learned-tracker arms. |
| `unseen_and_vehicle_picks_lock_files` | evaluation | m4a, m4b | files | 23 | 567,142 | `artifacts/traverse` | Top-level pick lock files and job lists of the unseen-arena, Gator and Polaris evaluations. |
| `unseen_arena_drive_folders_rigid` | evaluation | m4a | tar | 5 | 1,861,056,555 | `artifacts/traverse/arena_gator_20260925/e6/runs_rigid` | Drive folders behind the 37,100 cells of m4_unseen_arenas_hmmwv_rigid.csv (36,653 distinct drives). |
| `unseen_arena_drive_folders_soil` | evaluation | m4a, m4b | tar | 3 | 411,316,657 | `artifacts/traverse/arena_gator_20260925/e6/runs_soil` | Drive folders behind the 9,000 non-empty cells of m4_unseen_arenas_hmmwv_soil.csv (8,746 distinct drives). |
| `unseen_arena_drive_task_lists` | evaluation | m4a, m4b | files | 18 | 31,399,393 | `artifacts/traverse/arena_gator_20260925/e6/tasks` | Task lists that drove the HMMWV soil and rigid unseen-arena evaluations (including the Gator rigid f104 drives). |
| `unseen_arena_eval_indexes` | evaluation | m4a, m4b | files | 11 | 80,506,961 | `artifacts/traverse/arena_gator_20260925/e6` | Per-drive evaluation indexes of the unseen-arena study (soil and rigid; the rigid index also holds the Gator rigid f104 rows) and the task-to-arm mapping files. |
| `unseen_arena_route_picks` | evaluation | m4a, m4b | tar | 3 | 228,431,108 | `artifacts/traverse/arena_gator_20260925/e6/picks` | Locked planner picks of every HMMWV arm of the two unseen-arena tables, on the 8 unseen arenas, f104, the held-out and dev arenas (236 arm x arena folders). |
| `unseen_arena_suites` | evaluation | m4a, m4b | tar | 2 | 43,694,976 | `artifacts/traverse/arena_gator_20260925` | Frozen start/goal suites of the 8 unseen test arenas, held-out g203/g228 and dev g217: case files, designed routes, per-suite lock files, manifests, the f104 in-distribution list, the declared soil subset and the spread-headroom group list. |
| `vehicle_drive_task_lists` | evaluation | m4b | files | 16 | 6,562,883 | `artifacts/traverse` | Task lists of the Gator/HMMWV all-data drives, Polaris on f104, the gradient-refined Gator/HMMWV drives, Polaris on the unseen arenas and the smoke test. |
| `vehicle_eval_indexes_and_smoke_extract` | evaluation | m4b | files | 11 | 53,959,422 | `artifacts/traverse` | Per-drive indexes of the Gator all-data comparison, Polaris on f104 and Polaris on the unseen arenas, their task-mapping files, the per-route smoke and collection extract, and the definition of smoke sample A. |
| `vehicle_f104_drive_folders` | evaluation | m4b | tar | 3 | 702,007,891 | `artifacts/traverse` | Drive folders behind the 9,600 cells of m4_vehicles_f104_soil.csv (9,243 distinct drives; Gator, HMMWV and Polaris variants). |
| `vehicle_route_picks` | evaluation | m4b | tar | 3 | 86,180,277 | `artifacts/traverse` | Locked picks of the vehicle arms: Gator own models, HMMWV-trained model, gradient refinements, Polaris on f104 and on the 8 unseen arenas, and the Gator rigid planner. |
| `vehicle_smoke_test_drives` | evaluation | m4b | tar | 3 | 148,219,011 | `artifacts/traverse/offroad_vehicles_20260927/soil_v1` | Complete drive folders of the vehicle smoke test on sample A (144 f104 soil routes) for the 7 variants with their own folders (three Polaris driveline/wheel variants, stock and re-geared M113, Gator at 1 ms and 0.5 ms steps), plus the failure record of the one crashed drive. The polaris_stock and hmmwv_stored columns of m4_vehicle_smoke.csv come from drives released in the raw collections polaris_f104_soil_collection_runs and hmmwv_soil_f104_collection (not required: about 3.7 GB). |
|