l1ziang/LightWAM-AD-e100

LightWAM-AD: epoch 100

A one-pass driving policy adapted from Wan2.1-T2V-1.3B with future-video supervision and direct trajectory regression. Code and architecture.

Checkpoint and dependencies

step_131700.pt uses the custom lightwam_ad_v1 format. It contains LoRA, residual adapters, the trajectory head, the trained video output head and the ego-state projection. It is an adaptation checkpoint, not a standalone model or a standard Transformers/PEFT from_pretrained package. Optimizer state is not included.

The original Wan DiT, VAE, T5 and tokenizer are not redistributed here. Download them from the official Wan2.1-T2V-1.3B repository. DiT and VAE are needed for inference. T5 and its tokenizer prepare the static text-embedding cache; once the matching cache exists, inference does not load T5. NAVSIM data, maps and scoring caches must also be obtained separately.

Download this checkpoint:

hf download l1ziang/LightWAM-AD-e100 step_131700.pt SHA256SUMS --local-dir ./lightwam-ad-e100

Follow the setup, cache preparation and evaluation guide, using task=lightwam_ad_navsim_front_384x672 and CKPT=/path/to/lightwam-ad-e100/step_131700.pt. The fixed action normalization is implemented in the codebase.

Training and results

NAVSIM/OpenScene adaptation: 84,258 training and 851 validation clips; 100 epochs, effective batch 64, rank-64 LoRA and BF16. Training uses current and future front images with expert trajectories; no SimScale data or RL. Inference takes the current 384 × 672 front image, ego state and cached text, and outputs eight ego-frame poses over four seconds.

NAVSIM-v2 EPDMS (0–100)	Score
navtest	88.576
navhard, combined	33.281

On one H200, batch 1 and BF16: mean E2E latency is 71.570 ms, or 63.575 ms with the optional unused-tail skip. E2E includes in-memory preprocessing and transfers, but excludes disk loading and PDM scoring. Full-model parameters: 1,656,343,350 loaded and 110,555,459 trainable. These counts include the separately obtained base model, not just this weights file.

Scores use the baseline inference path. The optional skip has exact output parity on 200 sampled scenes, not a separate full NAVSIM evaluation. Full experiment record. This is a research checkpoint; the reported evaluations do not establish real-vehicle deployment readiness.

Attribution

We thank SimWAM for the codebase and training/evaluation infrastructure, and the Wan, Light-WAM, Fast-WAM and NAVSIM teams. The codebase retains its MIT license and upstream notices; the separately downloaded Wan assets retain their Apache-2.0 license.