Reproduce “LTLDiff: Finite Linear Temporal Logic-Guided Data Generation and Diffusion Policies for Multi-agent 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
Chuhan Meng, Haiyan Yin
LTLDiff combines finite-horizon temporal logic with diffusion-based control to generate and execute coordinated manipulation trajectories for multiple robots. By using the same LTLf specification for demonstration generation, policy conditioning, and joint trajectory guidance, it improves several one-, two-, and three-agent RoboFactory tasks over an unconstrained diffusion baseline, while remaining unreliable on long-horizon four-agent tasks.
Key insights
- Using temporal logic at both data-generation and policy-learning stages is more effective than applying it at only one stage; the full configuration outperforms generation-only and training-only ablations.
- Joint formula guidance can coordinate decentralized agents without requiring a single fully centralized policy, because agents share the LTLf embedding and receive corrections from a joint trajectory regressor.
- Logical conditioning helps most on structured coordination tasks, but it does not eliminate difficulties caused by visual manipulation, long horizons, sequential handoffs, or errors in language-model-generated formulas.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.