Answers are grounded in this paper and cite their sources.
We present a four-stage Learning from Demonstration (LfD) pipeline that incrementally constructs a general policy for multi-stage contact-rich manipulation from a single human demonstration. The pipeline is built around a crucial inductive bias—environmental constraints—that separates general task structure from hard-to-model instance-specific details. We show that, on top of this inductive bias, simple and well-established components are already sufficient to build an interactive LfD system that generalizes to unseen variations in object pose, local geometry, and contact conditions under real-world uncertainty. We validate the system on seven real-world multi-stage contact-rich manipulation tasks and achieve over 90% success across tasks. These results establish environmental constraints as fundamental building blocks for efficient generalization in learning from demonstration. The project website can be found at https://lfdec.github.io/ .
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: