RCFC — Lifelong Robot Imitation Learning

Prototype replay and coarse-to-fine compatibility for lifelong imitation learning.

RCFC is a lightweight framework for lifelong imitation learning that combines compact prototype replay with coarse-to-fine compatibility regularization to reduce catastrophic forgetting and improve cross-task transfer.

  • Maintains a small set of per-task prototypes instead of full trajectory buffers.
  • A two-stage compatibility loss aligns new policies with previous ones at both coarse (task-level) and fine (step-level) granularities.
  • Evaluated on the LIBERO benchmark across sequential task streams, with consistent gains over replay- and regularization-based baselines. Currently under review at ICML 2026.