InDRiVE: Reward-Free World-Model Pretraining for Autonomous Driving via Latent Disagreement

arXiv 25.12 2025 51.6 method, application

TLDR

InDRiVE uses latent disagreement for reward-free world-model pretraining in autonomous driving, enabling zero-shot and few-shot adaptation in CARLA.

Reasoning

The paper presents a novel reward-free pretraining method using latent ensemble disagreement as intrinsic motivation, which is a strength for reducing reward engineering. However, it is only evaluated in the CARLA simulator, lacking real-world validation, and the approach is specific to autonomous driving.

Read-first score

Read-first score 51.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 8%

86.7

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

Topical relevance 42%

58.6

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

Methodology quality 25%

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 466.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

generative world model

interactive world model

world simulator

video world model

Deep Analysis

Innovations

Reward-free pretraining using intrinsic motivation from latent ensemble disagreement as a proxy for epistemic uncertainty
DreamerV3-style MBRL agent for autonomous driving with reward-free pretraining
Imagination-based actor-critic learning a planner-free exploration policy directly from the learned world model
Zero-shot transfer by freezing all parameters and deploying pretrained exploration policy in unseen towns and routes
Few-shot adaptation with limited extrinsic feedback for downstream objectives (lane following and collision avoidance)

Methodology

The paper presents InDRiVE, a DreamerV3-style model-based reinforcement learning agent that performs reward-free pretraining in the CARLA simulator using intrinsic motivation derived from latent ensemble disagreement. Disagreement acts as a proxy for epistemic uncertainty, driving the agent toward under-explored driving situations. An imagination-based actor-critic learns a planner-free exploration policy directly from the learned world model. After pretraining, zero-shot transfer is evaluated by freezing all parameters and deploying the policy in unseen towns and routes, followed by few-shot adaptation using limited extrinsic feedback for lane following and collision avoidance.

Key Results

Disagreement-based pretraining yields stronger zero-shot robustness and robust few-shot collision avoidance under town shift and matched interaction budgets in CARLA experiments across towns, routes, and traffic densities.

Resources

Paper