Preprint: DemoStream Collects Real-World Robot Demos at 66 Per Hour Using Batched Task-and-Motion Planning

The DemoStream system can automatically collect robot manipulation demonstrations on physical hardware at an average rate of 66 per hour, significantly outpacing traditional methods that rely on constant human teleoperation or are specialized for single tasks.

Learning from demonstration typically requires massive datasets, but existing collection approaches either demand continuous human involvement or lack generality. While Task and Motion Planning (TAMP) systems can autonomously solve general-purpose tasks and generate demonstrations, the multi-modality of their solutions has historically made it difficult to train successful behavior cloning policies from them.

Building on recent work leveraging GPU acceleration for TAMP, this study introduces DemoStream. It employs batched TAMP sampling operations and a new algorithm for batched constraint satisfaction to explore many solutions in parallel, optimizing demonstration quality. Simulation experiments ablate different action spaces, cost functions, and batch sizes to analyze their effect on downstream policy success rates. Finally, the authors demonstrate that they can leverage TAMP's safety guarantees to directly collect real-world demonstrations, reporting an average rate of 66 demos per hour.

This is an arXiv preprint (submitted October 15, 2026). All performance metrics are self-reported by the authors and have not yet been independently reproduced. Specific hardware configurations, task complexities, and the quantitative improvement in addressing 'multi-modality' require full-text verification.