Adapting a World Model for Trajectory Following in a 3D Game

arXiv 25.4 2025 35.4 method, application

TLDR

Applies Inverse Dynamics Models with various encoders and policy heads for trajectory following in a 3D game, evaluating deviation and alignment strategies.

Reasoning

Strengths include real-world evaluation in a modern 3D game and systematic comparison of encoders and policy heads. Weaknesses are the narrow focus on trajectory following rather than a comprehensive world model, and the abstract does not detail world model adaptation despite the title.

Read-first score

Read-first score 35.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 2.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,result

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Topical relevance 42%

2.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 233.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Application of Inverse Dynamics Models (IDM) with different encoders and policy heads to trajectory following in a modern 3D video game (Bleeding Edge)
Investigation of future alignment strategies to address distribution shift caused by aleatoric uncertainty and agent imperfections

Methodology

The study applies Inverse Dynamics Models (IDM) with various encoders (trained from scratch, DINOv2) and policy heads (GPT-style, MLP-style) to trajectory following in the 3D game Bleeding Edge. It evaluates performance using trajectory deviation distance and first significant deviation point, comparing configurations across diverse data, low data, and fine-tuning settings.

Key Results

In a diverse data setting, a GPT-style policy head with an encoder trained from scratch performs best; in a low data regime, DINOv2 encoder with GPT-style policy head yields best results; when pre-trained on diverse data and fine-tuned for a specific behavior, both GPT-style and MLP-style policy heads achieve comparable results.

Limitations

Optimal configuration depends on the chosen setting, requiring task-specific tuning
Evaluation is limited to a single 3D game (Bleeding Edge), so generalizability to other environments is not established
The approach relies on Inverse Dynamics Models, which may not capture all aspects of trajectory following under severe distribution shifts

Resources

Paper