| --- |
| license: other |
| license_name: physicalai-av-derived-research-only |
| tags: |
| - tanitad |
| - autonomous-driving |
| - anchored-diffusion |
| - planner |
| - closed-loop |
| - fine-tune |
| 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-C base — E1b failure-gated closed-loop SFT (research artifact) |
|
|
| **This checkpoint is a pre-registered `BOUND` result, not a deployed model.** |
| It is published for reproducibility and because it was the last single-disk |
| artifact of the E1b/E1c experiment line. |
|
|
| - **Base:** `Sayood/tanitad-refc-base` (REF-C anchored-diffusion planner, |
| 104,191,577 params, 128 anchors, 2 denoise steps), step 29999. |
| - **Fine-tune:** 4000 steps, lr 2e-5, cosine, warmup 100, encoder **frozen** |
| (13,732,945 trainable / 90,458,632 frozen). Two interleaved objectives: |
| R2LPL-shaped anchor-score/traj supervision toward a logged-corridor *recovery* |
| demonstration at 3,537 mined recoverable pre-failure states, plus an |
| open-loop replay branch on the parity-train corpus. |
| - **Mining/replay corpus:** `physicalai-train-e438721ae894` (2376 episodes). |
| **Evaluation:** a byte-level-disjoint 44-episode held-out set. |
|
|
| ## Measured (paired episode-cluster bootstrap, B=2000, over the held-out episodes) |
|
|
| | metric (K=185, 18.5 s closed loop) | base | this ckpt | paired delta | |
| |---|---|---|---| |
| | corridor-departure, overall | 0.5877 | 0.1603 | −0.4274 [−0.5161, −0.3378] **separated** | |
| | corridor-departure, junction | 0.8414 | 0.4144 | −0.4270 [−0.6838, −0.1648] **separated** | |
| | peak abs XTE (m) | 38.94 | 3.04 | −35.90 [−49.33, −24.12] **separated** | |
| | OOD peak ratio (in-band check) | 1.2664 | 1.1339 | −0.1325 **separated (favourable)** | |
|
|
| | open-loop guardrail (held-out) | base | this ckpt | paired delta | |
| |---|---|---|---| |
| | ADE@2s (m) | 0.4747 | 0.6693 | **+0.1947 [+0.1415, +0.2522] separated WORSE** | |
| | anchor accuracy | 0.6815 | 0.6163 | **−0.0651 separated WORSE** | |
| | anchor traj L1 | 0.1775 | 0.2399 | **+0.0624 separated WORSE** | |
|
|
| **Verdict: `BOUND`.** The pre-registration committed that a closed-loop win |
| bought with a CI-separated open-loop regression is *not* a success. The |
| diagnosed cause: the forgetting guard was monitored on the corpus it replays, |
| so it could only ever report success. E1c re-runs this configuration with the |
| guard on held-out data and reports the full frontier. |
|
|
| ## Files |
| `ckpt.pt` (model + optimizer, step 3999) · `config.json` · `metrics.json` · |
| `train_log.jsonl` |
|
|
| **Not a safety claim.** The closed loop here is map/agent-free and measures |
| corridor-keeping/drift, not collision or off-road safety. |
|
|
