GrndCtrl: Grounding World Models via Self-Supervised Reward Alignment

arXiv 25.12 2025 67.5 method

TLDR

GrndCtrl aligns pretrained video world models with geometric rewards via GRPO, improving spatial coherence and navigation stability in outdoor environments.

Reasoning

The paper introduces a novel self-supervised reward alignment framework (RLWG) to address geometric grounding in video world models, a key limitation for navigation tasks. Strengths include a clear methodology using GRPO and multiple geometric rewards, but the evaluation is limited to outdoor environments without explicit real-world benchmarks, and the abstract lacks details on dataset scale or comparisons.

Read-first score

Read-first score 67.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

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=baseline,evaluation,metric

Topical relevance 42%

78.6

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

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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 165.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

interactive world model

Deep Analysis

Innovations

Reinforcement Learning with World Grounding (RLWG), a self-supervised post-training framework for aligning pretrained world models with physically verifiable structure
GrndCtrl, a reward-aligned adaptation method using Group Relative Policy Optimization (GRPO) to achieve geometric and perceptual grounding
Use of multiple verifiable rewards (pose cycle-consistency, depth reprojection, temporal coherence) analogous to RLVR in language models

Methodology

RLWG post-trains pretrained video world models via reinforcement learning with geometric and perceptual rewards (pose cycle-consistency, depth reprojection, temporal coherence). The GrndCtrl instantiation employs Group Relative Policy Optimization (GRPO) to adapt the model, analogous to reinforcement learning from verifiable feedback in LLMs. Evaluation is performed on outdoor embodied navigation tasks, comparing against supervised fine-tuning baselines.

Key Results

GrndCtrl achieves superior spatial coherence and navigation stability over supervised fine-tuning in outdoor environments, demonstrating that verifiable reward alignment bridges generative pretraining and grounded behavior.

Limitations

The abstract does not explicitly state limitations; potential concerns include reliance on carefully designed verifiable rewards that may not generalize to all environments or tasks
The method's effectiveness is demonstrated only in outdoor navigation, leaving indoor or dynamic scenes unaddressed
Post-training with multiple rewards may introduce computational overhead and sensitivity to reward weighting

Resources

Paper