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.
Trained on NVIDIA PhysicalAI-AV. Access granted per request for research/eval use only; no redistribution.
Log in or Sign Up to review the conditions and access this model content.
TanitAD — Flagship 4-Brain
CORRECTION (2026-07-16): the checkpoint currently in this repo is the OLD no-speed run (predictor action_dim 2, step 22000) that an ops-supervisor bug kept alive. The definitive flagship is the speed-input variant (action_dim 3, v0 as 3rd action channel), now resumed and training; this repo will be replaced with the speed checkpoint once it is trained. The 2.918 m @2s gate cited below was measured on the no-speed run — treat it as the no-speed baseline, not the definitive flagship.
Hierarchical latent world model for autonomous driving (TanitAD 3-arm study: flagship / REF-A / REF-B). 4-brain: trained ViT-12 encoder (9-ch, 256px) -> operative/tactical/strategic predictors + H15 imagination; SigReg anti-collapse (free_dims 64). Held-out grounded-rollout ADE@2s (no-speed run) = 2.918 m vs CV 0.825 m. Eval needs the tanitad stack + flagship4b config (WorldModel). Trained on PhysicalAI-AV derived features.
- Downloads last month
- 17