Unified Driving Tokens: Representation- and Geometry-Guided Discrete Tokenizer for Driving World Models and Planning
TLDR
A discrete tokenizer guided by DINO features and geometry for driving world models and planning, evaluated on NAVSIM.
Reasoning
The paper introduces a novel tokenizer that aligns with DINO features and incorporates geometric cues, showing improved reconstruction and planning. However, evaluation is limited to NAVSIM and lacks comparison to other tokenizers.
Read-first score
Read-first score 58.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.
Field roles
Rank sensitivity
Stability: volatile; rank range: 404.
Keyword Scores
Deep Analysis
Innovations
- Representation-guided and geometry-enhanced tokenizer that learns discrete tokens under joint supervision
- Alignment of discrete bottleneck with frozen DINO feature space through feature decoding
- Injection of geometric state-related cues via adjacent-frame depth and relative-pose supervision
- Stabilization of joint objectives with multi-codebook quantization
- Evaluation of learned tokens with lightweight planning readout and GPT-style next-token world model
Methodology
The tokenizer is trained with joint supervision: feature decoding aligns the discrete bottleneck with a frozen DINO feature space, while RGB reconstruction uses perceptual and adversarial losses. Adjacent-frame depth and relative-pose supervision inject geometric cues, and multi-codebook quantization stabilizes the joint objectives. The same learned tokens are evaluated on NAVSIM using a lightweight planning readout and a GPT-style next-token world model.
Key Results
Experiments on NAVSIM show improved reconstruction fidelity and representation consistency, competitive planning performance under a fixed decoder, and better generative quality under matched settings.