Ctrl-World: A Controllable Generative World Model for Robot Manipulation

arXiv 25.10 2025 71.1 method, application

TLDR

A controllable multi-view generative world model for robot manipulation, enabling policy evaluation and improvement via long-horizon consistent imagination.

Reasoning

The paper addresses a key challenge in building controllable world models for generalist robot policies, using a large real-world dataset and achieving long-horizon consistency. However, the abstract is cut off, and potential limitations include reliance on a single dataset and lack of real-world deployment validation.

Read-first score

Read-first score 71.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 62.

Topical relevance 42%

88.6

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

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=dataset,evaluation

Reproducibility 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 174.

Keyword Scores

world model

generative world model

interactive world model

video world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Controllable multi-view world model for evaluating and improving generalist robot policies
Pose-conditioned memory retrieval mechanism for long-horizon consistency
Frame-level action conditioning for precise action control

Methodology

The model is trained on the DROID dataset (95k trajectories, 564 scenes) and generates spatially and temporally consistent trajectories for over 20 seconds. It uses a pose-conditioned memory retrieval mechanism to maintain long-horizon consistency and frame-level action conditioning to achieve precise action control, enabling multi-step interactions with generalist robot policies.

Key Results

The method can accurately rank policy performance without real-world robot rollouts, and by synthesizing successful trajectories in imagination for supervised fine-tuning, it improves policy success by 44.7%.

Limitations

Evaluation is limited to the DROID dataset, which may not cover all real-world scenarios
Long-horizon consistency is demonstrated for over 20 seconds, but longer interactions may not be supported
The approach requires a large dataset (95k trajectories) and may not generalize to novel camera placements or scenes beyond those in the training data

Resources

Paper