Reproduce “Temporal Logic Guidance for Action-Only Diffusion Policies with World Models” 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
Moritz Zoellner, Anastasios Manganaris, Rohan Paleja
A method enables action-only diffusion policies to satisfy temporal logic constraints at inference time by integrating a separately trained, differentiable world model for state prediction. This approach reduced constraint violations from 84% to 4% while maintaining task success in a can transport task, ensuring objects like cans remain upright.
Key insights
- Decoupling action generation from state prediction, through a differentiable world model, allows state-dependent formal guidance (like STL) to be applied to a wider range of modern action-only robot policies.
- Gradient-based guidance within the diffusion denoising process is significantly more effective at enforcing complex temporal constraints than simple sample-and-rank baseline methods.
- This inference-time guidance can drastically improve adherence to specified constraints (e.g., safety, preferences) without compromising the robot's ability to successfully complete the primary task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.