Hugging Face's logo

Sayood
/
tanitad-refc-xl

tanitad
autonomous-driving
anchored-diffusion
trajectory-prediction
planner
reference-arm
Model card Files
xet
 Community

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These weights are trained on NVIDIA PhysicalAI-AV data (TanitAD research program). Access is granted per request for research/evaluation use only; you agree not to redistribute.

Log in or Sign Up to review the conditions and access this model content.

TanitAD β€” REF-C-XL (anchored diffusion, step 29,999 / 30,000 FINAL)

Reference arm C of the TanitAD program: a DiffusionDrive-style anchored-diffusion direct planner β€” the budget-matched non-world-model control for the hierarchical 4-brain flagship. This is the XL rung (251.9 M params) of the three-size REF-C ladder (small 54.7 M Β· base 104.2 M Β· XL 251.9 M).

Registry key: refc-diffusion-xl-30k Β· TanitEval arm key: refc-xl-30k

Source of truth for every number below: Project Steering/MODEL_REGISTRY.md Β§4.1 and the raw eval JSON taniteval/results/refc-xl-30k.json. Evidence class: MEASURED unless marked otherwise.

Architecture

Read from this run's own config.json (shipped in this repo).

  • Encoder β€” torchvision-free ResNet-L, in_channels 9 (3 RGB frames at 100 ms spacing, channel-stacked), image_size 256, base_width 124, blocks (3, 8, 20, 6).
  • Anchor vocabulary β€” 256 trajectory anchors built by furthest-point sampling from a 4,096-window pool, seed 0 (FPS, not k-means: the corpus is ~74 % straight and k-means collapses onto the straight mode).
  • Decoder β€” anchor queries cross-attend the conv feature map: d 512, 8 heads, 6 layers, ff_mult 4, aux_hidden 512, 2 truncated-denoise steps, noise_std 0.1. Emits per-anchor confidence + offset.
  • Measurement encoder β€” hidden 128, d_out 128, ego_dropout 0.5 (anti-shortcut).
  • LAW latent-world-model auxiliary β€” hidden 2048.
  • Strategic-context graft β€” hidden 768, d_ctx 96; hierarchy true, graft_maneuver true (H19 maneuver-to-anchor prior, live from step 0).
  • Imagination graft (H15 belief field) β€” ON, XL-only: d 512, depth 6, 8 heads, ff_mult 4, head_hidden 1024.
  • Horizons (5, 10, 15, 20) steps @ 10 Hz = 0.5 / 1 / 1.5 / 2 s Β· path_dists (2, 5, 10, 20) m Β· speed_bins 4, speed_max 30.0 Β· graft_target_latent false, grounded_selector false, refc1 false.

Parameters β€” measured at instantiation (config.json param_breakdown)

module params
encoder 199,496,532
decoder 22,702,345
imagination (H15) 20,986,339
strategic 4,133,472
law 4,082,656
aux 513,960
measurement 17,280
total 251,932,584

n_params_trainable = 251,932,584 (all trainable β€” no frozen encoder). Anchors are buffers, not parameters (~0.048 M of buffers total).

Training

item value
Corpus NVIDIA PhysicalAI-AV front-wide, 2,376 episodes / 406,099 windows (train split)
Strict-parity build key physicalai-train-e438721ae894
Corrupt-clip skip-hash f09e44db (24 corrupt front-wide clips excluded)
Steps 29,999 / 30,000 (metrics.json final.step = 29999, steps = 30000)
Optimizer Adam (DiffusionDrive/TCP convention, not AdamW), lr 1e-4, warmup 2000, cosine
Loss weights traj 1.0 Β· cls 1.0 Β· law 0.5 Β· route 0.1 Β· man 0.1 Β· speed_cls 0.2
Batch / workers 20 / 6 Β· seed 0 Β· device cuda
Route labels v1 (route_target(nav_cmd)) β€” see caveat 5
Hardware tanitad-pod3; finished 2026-07-20 09:19 UTC

Exact command (read from live ps on the training pod):

cd /workspace/TanitAD/stack && PYTHONPATH=/workspace/TanitAD/stack \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True nohup python3 scripts/refc_train.py \
  --data-root /workspace/pai_epcache \
  --out /workspace/experiments/refc-diffusion-xl-30k \
  --steps 30000 --mode diffusion --config xl \
  --anchors /workspace/experiments/refc_anchors_full.pt \
  --batch 20 --workers 6

Code: stack/tanitad/refs/refc.py + stack/scripts/refc_train.py (commits 6025769 redesign, 7e9c402 sizing; 15 tests/test_refc.py pass). Anchors built by stack/scripts/build_refc_anchors.py.

Final training metrics (step 29,999, metrics.json): loss 1.86033 Β· traj 0.17534 Β· cls 1.50755 Β· law 0.02031 Β· route 0.91015 Β· man 0.76259 Β· anchor_acc 0.50 Β· man_acc 0.75.

Evaluation

TanitEval taniteval.refc_eval, open-loop, on the clean held-out split physicalai-val-0c5f7dac3b11 β€” 40 episodes β†’ 881 windows, episode-disjoint from train. Protocol: window 8, stride 8, K = 20 steps @ 10 Hz, waypoints [5, 10, 15, 20], metric-BEV ego frame, nav = follow, decode = anchored diffusion with 2 truncated-denoise steps over 256 anchors, argmax-confidence select. ckpt_step recorded in the result JSON = 29999.

Headline β€” decision-grade (n = 881 windows / 40 episodes)

metric value estimator
ADE@2s (full-set) 0.4714 Β· CI95 [0.3896, 0.5556] (Β±0.0830) episode-cluster bootstrap, B = 2000 over the 40 val episodes (taniteval/ci.py)
FDE@2s (full-set) 1.0061 Β· [0.8301, 1.1875] same
miss@2m (full-set) 0.1419 Β· [0.0943, 0.1918] same
TMS-openloop (full-set) 0.2135 point estimate
ADE@0.5 / 1 / 1.5 s (full-set) 0.0681 / 0.1592 / 0.2932 point estimates

Trivial floor on the same split: constant-velocity ADE@2s 0.8377 (full-set) / 0.8248 (heldout). REF-C-XL beats CV overall and in every stratum.

Estimator caveat β€” never quote an interval without its estimator. An earlier published row read 0.458 Β± 0.057. That Β± is overlapping_holdout_se (historically mislabelled "8-split episode-disjoint jackknife"); it is neither a jackknife nor a valid SE and is measured 1.28–2.06Γ— too narrow across 10 arms (1.45Γ— for this arm). It is retained here only for continuity with older publications. Use the episode-cluster bootstrap row. Source: Project Steering/CI_RECOMPUTE_2026-07-20.json.

Strata (full-set, step 29,999)

stratum REF-C-XL ADE@2s CV baseline n
speed high 0.3243 0.6468 294
speed med 0.4989 0.9345 293
speed low 0.5912 0.9322 294
curv straight 0.3865 0.4393 634
curv gentle 0.6751 1.3566 125
curv sharp 0.7040 2.3764 122

The high-speed stratum is where the hierarchical flagship is weakest (0.5513) β€” a direct-head diffusion arm beats the world-model stack there by β‰ˆ 41 %.

Efficiency β€” batch 1, one A40, identical precision flags

fp32 tf32 amp16
plan tick p50 44.06 ms 27.78 ms 21.00 ms
plan tick p99 (fp32) 44.44 ms β€” β€”
GFLOPs / peak alloc (fp32) 702.2 / 1178.4 MB β€” β€”

Encoder is 88.7 % of the tick. Meets a 10 Hz budget in all three precisions. (REF-C-base is 1.32–2.02Γ— faster at statistically indistinguishable accuracy β€” see caveat 2.)

Honest caveats β€” all recorded in the registry; read them before quoting this model

  1. Flagship v1 vs REF-C-XL is a TIE, not an ordering. Under the paired episode-cluster bootstrap, Ξ”(REF-C βˆ’ flagship v1) = +0.0443 m, CI95 [βˆ’0.0544, +0.1465], P(Ξ”>0) = 0.809 β†’ NOT separated. A prior claim that REF-C finished "0.006 m behind the flagship" was RETRACTED: 0.006 m was a difference of split means; the full-set gap is 0.0443 m (8Γ— larger) and still not separated. Per-window correlation is only 0.207 here, so pairing buys little power β€” the tie is real, not a weak test.

  2. REF-C-base (104.2 M) ties this model on everything that ships. Paired Ξ”(base βˆ’ XL): ADE@2s +0.0013 [βˆ’0.0281, +0.0316], FDE βˆ’0.0030 [βˆ’0.0619, +0.0584], miss +0.0000 [βˆ’0.0261, +0.0272] β€” all NOT separated, per-window correlation 0.789 (the test is not weak). A 2.42Γ— parameter cut and a 2.20Γ— encoder cut cost nothing measurable on this corpus.

  3. Selection flaw β€” REF-C ranks with the UN-refined anchor's score. All 256 anchors are denoised, but selection uses the t=0 classifier score over the original anchors; the denoise passes' own confidences are discarded. Geometry is refined, ranking is not. Corpus figures (n = 881): selected 0.4714 Β· oracle-in-fan 0.1640 Β· gap 0.3075 m Β· frac_sel_2x_worse 0.454.

  4. The oracle gap is ~92 % IRREDUCIBLE β€” it is not available headroom. Settled across 47 trained arms: a learned re-scorer recovers at most 8.4 % of it on its own training data; a hand-written cost re-rank recovers 0.0 %; scoring the refined confidences is 2.9Γ— WORSE than baseline (that head is unsupervised at denoise timesteps). Do not add a target-speed term to the selection score (REFUTED: a GT-perfect speed matcher scores worse than baseline). Selection is not the productive lever on this architecture.

  5. Route-label confound vs the base and small rungs. XL trained with v1 route labels (route_target(nav_cmd) β€” circular and straight-by-default; labels_v2 was never set in refc_train.py). REF-C-base and REF-C-small trained with v2.1. XL-vs-base therefore conflates scale and labels. Calibration (flagship v1.5, the only end-to-end measurement of the label change): ADE +0.025 m, not CI-separated, but oracle βˆ’0.058 m. Do not present a clean scaling conclusion. The clean resolution is one control run (XL-with-v2.1 or base-with-v1) β€” not yet run.

  6. The fan lever is anchor-vocabulary WIDTH, not encoder scale. base's 128 anchors are a bit-exact prefix of XL's 256, so the fans compare over an identical vocabulary. Oracle-in-fan over the first K: base ≀ XL at every matched K (4β†’128); XL's entire oracle advantage arrives with anchors 129–256. Anchors are buffers (~0.048 M), so widening the vocabulary is nearly free, while widening the encoder demonstrably bought nothing here. (Caveat: a prefix restriction structurally penalises XL, whose winner-takes-all training spread modes across 256 slots.)

  7. Closed-loop numbers are ENV-CONFOUNDED β€” do NOT read them as a model result. On the AlpaSim NuRec suite (n = 12) this arm scores at-fault collision 33.3 %, off-road 25.0 %, pass 5/12, mean score 0.246. But REF-C's open-loop ADE on those same reconstructions is 1.52 m β€” 3.21Γ— its real-footage 0.4714, i.e. the input is ~3Γ— off the training distribution. Those rates measure model Γ— reconstruction fidelity, not the model. n = 12 also means one scene = 8.3 pp.

  8. Reproducibility gap (declared, not hidden): the TanitEval harness that produced every number above lives on tanitad-eval:/root/taniteval and is not committed to the TanitAD repo. The checkpoint survives; the exact evaluator currently does not travel with it.

  9. This is a research reference arm, not a deployable driving system. Open-loop ADE does not predict closed-loop behaviour (measured elsewhere in this program: 0.45 m open-loop β†’ 1.69 m closed-loop for the flagship). Claim strength on all metrics above: open-loop / weak.

Files

file contents
ckpt.pt full training checkpoint β€” {model: state_dict (673 tensors), opt: optimizer state, step: 29999}, 3,024,021,445 B, md5 966d4eff1ea5ddf86efba01b8344e198
config.json the run's own config: architecture, args, optimizer, loss weights, param_breakdown
metrics.json final training + val metrics at step 29,999

Verify after download: md5sum ckpt.pt must print 966d4eff1ea5ddf86efba01b8344e198.

Intended use / license

Internal research and evaluation only. Derived from NVIDIA PhysicalAI-AV (gated corpus) β€” not redistributable. Not a deployable driving system.

Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support