1367-GRM Replacing Stored Memory with Generative Replay to Reduce Catastrophic Forgetting in Lifelong Robot Navigation
Neural network navigation policies overwrite old knowledge when trained on new data, a problem called catastrophic forgetting. A robot that adapts to a new environment can lose earlier skills, including safety-critical ones like passing through narrow gaps. Lifelong Learning for Navigation (LLfN) and C-NAV deal with this by storing past experiences. However, their memory needs grow with each new environment, which is a problem for robots with limited onboard storage. We propose generative replay: a single fixed-size generator recreates representative and rare past situations on demand. The goal is to preserve navigation skills across many environments without memory growing.
Authors
- Anamo Kisho
- Mithil Honkanadavar
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- article
- Field-Weighted Citation Impact
- 0.00