Orbis: Overcoming Challenges of Long-Horizon Prediction in Driving World Models
TLDR
A driving world model using simple design choices achieves state-of-the-art long-horizon prediction, favoring continuous autoregressive over discrete token models.
Reasoning
The paper's strengths include a simple yet effective design, no need for extra supervision, and a fair comparison between discrete and continuous token models. Weaknesses are the limited training data (280h) and lack of exploration into interactive or reinforcement learning settings.
Read-first score
Read-first score 71.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.
Field roles
Rank sensitivity
Stability: volatile; rank range: 90.
Keyword Scores
Deep Analysis
Innovations
- Simple design choices without additional supervision or sensors (maps, depth, multiple cameras) achieving state-of-the-art performance
- Hybrid tokenizer compatible with both discrete and continuous approaches enabling fair side-by-side comparison
- Demonstration that continuous autoregressive models are less brittle and more powerful than discrete token models for driving world models
Methodology
The authors develop a world model using simple design choices, trained on 280 hours of video data with 469M parameters. They design a hybrid tokenizer that supports both discrete token models (based on flow matching) and continuous autoregressive models, allowing a direct comparison. Evaluation focuses on long-horizon prediction and generalization, particularly in turning maneuvers and urban traffic scenarios.
Key Results
The model achieves state-of-the-art performance despite its smaller size and limited training data, excelling in difficult scenarios like turning and urban traffic. The continuous autoregressive model outperforms the discrete token model, being less sensitive to design choices and more powerful.