README.md

---
license: other
license_name: physicalai-av-derived-research-only
tags:
- tanitad
- world-model
- autonomous-driving
- planner
extra_gated_prompt: >-
  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.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
---

# TanitAD — REF-B v2 (arch-v2 --refbpatch, step 29,999 / 30,000 FINAL)

Reference arm B of the TanitAD 3-arm study: a **hierarchical planner with a from-scratch trained encoder** (the "trained-encoder" reference vs REF-A's frozen encoder). This is the **v2** milestone-final checkpoint (271.6 M params).

## Architecture (v2)
- **Trained ViT encoder**: 9-channel, 256 px, patch 16, **depth 25**, d_model 768 (grad-checkpointed).
- **operative** (d 768, depth 6, action_seq 5) + **tactical** + **strategic** (4 nav commands) + confidence head + OOD monitor.
- **v2 --refbpatch levers**:
  - **B1 — TIME-anchored tactical decoder** (anchor space=time, 128 anchors from a 4096 pool, 4 layers / 8 heads) producing waypoints @ 5/10/15/20 steps (0.5/1/1.5/2 s).
  - **B2 — proprioceptive ego conditioning** `[v0, yr0]`: ego speed + backward-diff yaw-rate (raw rad/s), added to the nav embedding with **ego_dropout 0.5** (anti-shortcut).
  - **aux-yaw + aux-accel** heads, **path heads** at fixed distances [2, 5, 10, 20] m, jerk penalty 0.02.

## Training
- Data: **PhysicalAI-AV** front-wide (phase-0 epcache).
- **Step 29,999 / 30,000 (FINAL).** Optimizer AdamW (lr 3e-4, cosine), weight_decay 0.05, warmup 2000, AMP.

## Evaluation (TanitEval, held-out PhysicalAI-AV val, n=881 windows; open-loop)
Planner-native trajectory = tactical waypoint heads (direct regression, nav=follow). Bootstrap 95% CI.

| horizon | ADE (m) | CV baseline |
|--------|---------|-------------|
| 0.5 s  | 0.103   |             |
| 1.0 s  | 0.217   |             |
| 1.5 s  | 0.379   |             |
| 2.0 s  | **0.592** ± 0.069 | 0.825 |

- **fde@2s 1.23** (CV 1.71), **miss@2m 0.203** (CV 0.313). **Beats the constant-velocity baseline** (0.592 vs 0.825, −28%).
- **By speed** (model / CV, ADE@2s): high 0.43 / 0.65 · med 0.52 / 0.93 · low 0.82 / 0.93 — beats CV in every band.
- **By curvature** (model / CV, ADE@2s): gentle 0.86 / 1.36 · **sharp 0.90 / 2.38** (large margin) · straight 0.48 / 0.44 (CV is near-optimal on straights).
- **Error is ~79 % longitudinal** at 2 s (along-track / speed-timing dominant); becomes lateral-dominant only on sharp curves.
- **Vision use**: zeroing the camera frames worsens ADE@2s 0.591 → 0.965 (**≈39 % of accuracy is vision-driven**); the remainder rides the `[v0, yr0]` ego state.

### Milestone / cross-arm A/B (paired, same windows)
- **vs REF-B v2 @20k** (prior milestone): 0.591 vs 0.643 — **v2-30k wins, significant** (ΔCI [−0.077, −0.026]); largest gain on sharp curves (−0.135 m).
- **vs REF-B v1** (step-6k, ADE@2s 0.868): **v2-30k wins decisively** (ΔCI [−0.32, −0.22]); sharp-curve gain −1.19 m.
- **vs Flagship-4B v1 @30k** (grounded world-model): flagship better on open-loop ADE (0.427 vs 0.591).

### Generalization (cross-corpus, planner-direct)
| corpus | kind | ADE@2s | CV |
|--------|------|--------|----|
| PhysicalAI-AV | in-dist | 0.592 | 0.825 |
| comma2k19 | OOD (real highway) | 0.515 | 0.646 |
| Cosmos-Drive-Dreams | OOD synthetic | 0.495 | 0.451 |

The edge over CV shrinks OOD (in-dist −28 % → comma −20 % → cosmos +10 %, i.e. below CV). **Cosmos numbers are PROVISIONAL** (a subset of clips carry a ground-truth timing desync; a rebuilt cache is pending). Absolute OOD ADE is lower mainly because those corpora are geometrically easier (highway / synthetic, low curvature).

_Imagination and hierarchy panels are N/A for REF-B (a direct planner with no grounded rollout head). Trained on PhysicalAI-AV derived data. Gated for research/eval use._