Foresight: Failure Detection for Long-Horizon Robotic Manipulation with Action-Conditioned World Model Latents

arXiv 2026 53.4 method, application

TLDR

Foresight detects failures in long-horizon robotic manipulation using action-conditioned world model latents and conformal prediction, validated in simulation and real robots.

Reasoning

The paper addresses an underexplored problem with a novel framework that leverages world model embeddings for failure detection, requiring only final task labels. Strengths include real-world validation and adaptive threshold calibration; weaknesses are limited detail on limitations and comparison baselines in the abstract.

Read-first score

Read-first score 53.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 27.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

95.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95172744

Methodology quality 18%

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

Topical relevance 29%

38.6

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 18%

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

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 387.

Keyword Scores

world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Using action-conditioned world model latents as a unified representation for failure detection across different policies
Training failure detection with only final task-level success/failure labels, eliminating need for dense temporal annotations
Adaptive threshold calibration via functional conformal prediction (FCP) for robust detection

Methodology

Foresight trains an action-conditioned world model to produce latent embeddings from manipulation trajectories. These embeddings are used as input to a failure detector that is trained solely on final task-level success/failure labels. Detection thresholds are adaptively calibrated using functional conformal prediction to handle varying conditions.

Key Results

Foresight outperforms state-of-the-art failure detection methods across multiple simulation benchmarks (LIBERO-Long, ManiSkill-Long, BEHAVIOR-1K) and is validated on real robots (ReactorX-200 arm and Franka arm) on long-horizon tasks.

Limitations

Requires training an action-conditioned world model, which may be computationally expensive and data-intensive
Performance may depend on the quality and coverage of the world model embeddings
Evaluation is limited to specific simulation environments and a small set of real-robot tasks

Resources

Paper