Enhancing Physical Consistency in Lightweight World Models

2025-09-15

AI Summary

To address the trade-off between weak physical consistency in lightweight world models and the deployment difficulty of large models, this paper proposes a compact BEV-based world model. Methodologically: (1) a soft-mask training mechanism is designed to enhance physical interaction modeling for dynamic objects; (2) a zero-shot-compatible warm-start inference strategy is introduced to improve prediction stability and convergence efficiency. Experiments show that, at equal parameter counts, our method achieves a 60.6% higher weighted composite performance over baselines; even the smallest variant (130M parameters) outperforms baselines by 7.4%, with a 28% inference speedup. The core contribution lies in the first integration of soft masking and warm-start inference into a BEV world model framework—enabling substantial computational savings while preserving high-fidelity physical dynamics modeling.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Machine Learning: Hardware-aware ML

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

A major challenge in deploying world models is the trade-off between size and performance. Large world models can capture rich physical dynamics but require massive computing resources, making them impractical for edge devices. Small world models are easier to deploy but often struggle to learn accurate physics, leading to poor predictions. We propose the Physics-Informed BEV World Model (PIWM), a compact model designed to efficiently capture physical interactions in bird's-eye-view (BEV) representations. PIWM uses Soft Mask during training to improve dynamic object modeling and future prediction. We also introduce a simple yet effective technique, Warm Start, for inference to enhance prediction quality with a zero-shot model. Experiments show that at the same parameter scale (400M), PIWM surpasses the baseline by 60.6% in weighted overall score. Moreover, even when compared with the largest baseline model (400M), the smallest PIWM (130M Soft Mask) achieves a 7.4% higher weighted overall score with a 28% faster inference speed.

Problem

• Balancing model size and physical accuracy
• Improving lightweight world model physics consistency
• Enhancing edge device deployment with efficient predictions

Innovation

• Physics-Informed BEV World Model
• Soft Mask training technique
• Warm Start inference method