VISTAv2: World Imagination for Indoor Vision-and-Language Navigation

arXiv 25.11 2025 68.5 method, application

TLDR

VISTAv2 uses a generative world model to imagine future views for indoor vision-and-language navigation, improving planning via an online value map.

Reasoning

The paper introduces a novel integration of generative video prediction with value-based planning, demonstrating strong empirical results on MP3D and RoboTHOR. However, it is limited to short-horizon futures and evaluated on only two datasets, leaving scalability and long-horizon generalization unaddressed.

Read-first score

Read-first score 68.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Topical relevance 42%

87.1

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

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=ablation,baseline

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: 258.

Keyword Scores

world model

generative world model

video world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

Action-conditioned egocentric future view prediction using a generative world model
Online value map for planning fused at score level with base objective
Action-aware Conditional Diffusion Transformer video predictor for synthesizing short-horizon futures
Differentiable imagination-to-value head that fuses multiple rollouts into an egocentric value map
Efficient rollouts in VAE latent space with distilled sampler and sparse decoding on a single consumer GPU
Instruction-guided value fusion via vision-language scorer

Methodology

VISTAv2 is a generative world model that rolls out egocentric future views conditioned on past observations, candidate action sequences, and instructions. It uses an action-aware Conditional Diffusion Transformer video predictor to synthesize short-horizon futures, aligns them with language via a vision-language scorer, and fuses multiple rollouts in a differentiable imagination-to-value head to output an egocentric value map. Rollouts are performed in VAE latent space with a distilled sampler and sparse decoding for efficiency. The model is evaluated on MP3D and RoboTHOR against strong baselines.

Key Results

VISTAv2 improves over strong baselines on MP3D and RoboTHOR, and ablations confirm that action-conditioned imagination, instruction-guided value fusion, and the online value-map planner are all critical to performance.

Limitations

Only evaluated on indoor environments (MP3D and RoboTHOR)
Requires candidate action sequences for rollouts, which may limit applicability in unknown or dynamic settings
Relies on short-horizon future predictions, potentially limiting long-horizon planning

Resources

Paper