Reproduce “STORM: Search-Guided Generative World Models for Robotic Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenjun Lin, Jensen Zhang, Kaitong Cai, Keze Wang
STORM introduces a framework for robotic manipulation that integrates diffusion-based action generation, a generative video world model, and Monte Carlo Tree Search for foresight-driven planning. This approach achieved an average success rate of 51.0% across challenging manipulation tasks, improving over the CogACT baseline by 3.1 percentage points and demonstrating robust recovery from initial failures.
Key insights
- Integrating visual foresight through a generative video world model significantly enhances robotic decision-making, allowing agents to "predict before they act" and overcome the limitations of purely reactive policies.
- Augmenting the generative world model with reward prediction is critical for learning task-relevant causal dynamics, transforming it into a "value-aware simulator" essential for effective MCTS planning.
- Search-guided planning (MCTS) with diverse action proposals enables robust recovery from initial failures by exploring alternative trajectories and adapting plans, a capability largely absent in reactive VLA models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.