TanitAD β rollout-recovery arms (RR-20 / RR-CTL)
Weights from the rollout-recovery experiment, pushed 2026-08-02. These existed on a single pod disk and nowhere else.
The experiment
Both arms fine-tune v1 (flagship4b-speedjerk-30k) for 2,000 steps to step 31999 β same pod,
same seed, exactly one flag different.
RR-CTL (--rollout-k 4) |
RR-20 (--rollout-k 20) |
|
|---|---|---|
| ade@2s | 0.424 [0.368, 0.483] | 0.348 [0.291, 0.411] |
| paired ΞCI | [0.0613, 0.0906] β separated, RR-20 wins | |
| speed bias | +0.9397 m/s | β -0.0092 m/s |
| speed MAE | 2.3122 | 1.6965 |
| along-track final bias | +0.3355 m | -0.0484 m |
| curvature MAE | 0.021798 1/m | β οΈ 0.048254 1/m |
| yaw-rate MAE | 8.848 Β°/s | 9.8445 Β°/s |
| miss@2m | 0.043 | 0.056 |
n = 881 windows over the canonical 40 validation episodes; paired episode-cluster bootstrap, B = 2000.
What it means
Training on longer recursive rollouts essentially ERASED the longitudinal speed bias (+0.94 β -0.009 m/s). That bias is the programme's largest measured defect β an oracle target-speed recovers 88.7 % of the goal-conditioning gap, all of it longitudinal.
β οΈ It was paid for in lateral fidelity: curvature error 2.2Γ worse, yaw-rate worse, and miss@2m rose while ADE fell. On ADE alone this reads as a clean win; it is a trade, visible only because the evaluation reports longitudinal, lateral, tactical and strategic families separately rather than a single scalar.
β RR-20 must be compared to RR-CTL, never to v1. Against v1 you would confound rollout-k
with 2,000 extra fine-tuning steps.
Contents
rr20/model_grounding.ptβ RR-20:model+grounding+step(optimizer state dropped)rrctl/model_grounding.ptβ RR-CTL, same structurerefc-base-e1f-junction/β a REF-C junction arm, also single-diskprovenance/β run status files
Trained on NVIDIA PhysicalAI-AV. Architecture flagship4b, 276.9 M params, speed_input on
(action_dim 3).