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.

TanitAD internal research checkpoint. Access requires owner approval.

TanitAD Flagship-4B — Phase-0 speed+jerk v1 (30k)

Hierarchical 4-brain latent world model for autonomous driving (~260M model params, 273M with grounding heads). Phase-0 reset arm ("speedjerk"): v0 speed input + jerk-smoothness + aux-accel grounding, trained 30k steps (batch 16, accum 4, lr 3e-4 cosine, bf16, rollout-k 4) on the canonical 2,376-episode PhysicalAI-AV front-cam epcache (key e438721ae894, f-theta v2 calibration).

Architecture

End-to-end trained ViT encoder (depth 12, 256px, 9ch stack) — LeWM-style
Action-conditioned operative predictor (FiLM, JEPA multi-horizon 1/2/4)
Trained tactical policy (maneuver vocab + goal latent + 2s sub-waypoints)
Trained strategic transformer (route/nav ctx, 15-25s horizon)
SIGReg (LeJEPA sliced Epps-Pulley normality) rank stabilization, free-dims 64
Grounding heads (metric inverse-dyn + SE(2) step readout) for metric decode

Final training metrics (step 29,999)

grounded operative fwd ADE: 0.033 m (per-step readout, k=4)
grounded tactical mid DE: 4.24 m (unimodal wp head — known weakness, v2 replaces it with a time-anchored multi-anchor decoder)
TanitEval @19k snapshot: open-loop ADE 0.615 m vs CTRV oracle 0.544 m / best-of-3 kinematic floor 0.50 m; skill-vs-floor on straights 1.49x (near-trivial-competitive on a 74%-straight corpus — honest caveat)
maneuver acc 0.94, route acc 1.0, erank 17.8

Files

ckpt.pt — full training checkpoint (model + grounding heads + optimizer)
config.json — the run config

Data / intended use

Internal research only (PhysicalAI-AV derived; gated manual access). Not a deployable driving system.

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support