Toward Physically Consistent Driving Video World Models under Challenging Trajectories
TLDR
Proposes PhyGenesis, a driving video world model that generates physically consistent videos under challenging trajectories using a physical condition generator and physics-enhanced video generator.
Reasoning
Strengths: Addresses a key limitation of existing world models (failure on counterfactual trajectories) with a novel two-component framework and a heterogeneous dataset. Weaknesses: Relies on CARLA simulator for supervision, which may not fully capture real-world physics; limited to driving domain.
Read-first score
Read-first score 64.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 208.
Keyword Scores
Deep Analysis
Innovations
- Physical condition generator that transforms invalid trajectory inputs into physically plausible conditions
- Physics-enhanced video generator for high-fidelity multi-view driving video generation
- Challenging-trajectory learning strategy for trajectory correction and physically consistent video generation
- Large-scale physics-rich heterogeneous dataset combining real-world data and CARLA-simulated challenging scenarios
Methodology
PhyGenesis comprises two key components: a physical condition generator that converts potentially invalid trajectory inputs into physically plausible conditions, and a physics-enhanced video generator that produces high-fidelity multi-view driving videos under these conditions. Training leverages a large-scale heterogeneous dataset that includes real-world driving videos and diverse challenging scenarios generated by the CARLA simulator, with supervision signals that guide the model to learn physically grounded dynamics under extreme conditions.
Key Results
Extensive experiments show that PhyGenesis consistently outperforms state-of-the-art methods, particularly on challenging trajectories.