Learning from demonstration is a common strategy for teaching robots manipulation skills, but it typically requires training on massive datasets of demonstrations. The new arXiv paper "DemoStream" proposes a batched task-and-motion-plan optimization approach to tackle this bottleneck.
The abstract states that existing demonstration collection approaches are insufficient, though the description is cut off in the source. Based on the title, DemoStream appears to jointly optimize task planning and motion planning in batches, potentially automating the generation of demonstrations rather than relying on manual collection.
The paper is listed on arXiv under identifier 2610.22020. Because the abstract is truncated, specific details about the method and experimental results are not available from this source alone.