Authors
Peng Qin, Jianjian Zhao, Jiayu Zhao, Tao Zhao
Published in
IEEE transactions on neural networks and learning systems. Volume PP. Nov 12, 2026. Epub Nov 12, 2026.
Abstract
In intelligent transportation systems, mobile robots must continuously adapt and navigate across diverse scenarios. However, most reinforcement learning (RL) methods assume a fixed task distribution, making them prone to catastrophic forgetting when facing incremental tasks. Although parameter consolidation techniques mitigate forgetting, their inherent bias limits cross-scenario navigation performance. To mitigate these limitations, a knowledge-guided continual RL method with integrative consolidation and memory augmentation (GCRL-ICMA) is proposed to alleviate key issues in multiscenario navigation, including forgetting, negative transfer, and low task learning efficiency. Specifically, for parameter consolidation, a complementary parameter importance estimation method is proposed by combining elastic weight consolidation (EWC) and synaptic intelligence (SI) to preserve knowledge from previous tasks. For memory enhancement, a dual experience replay mechanism is introduced, which mixes current experiences with clustered successful experiences from historical tasks to improve policy updates and data efficiency. In addition, a knowledge-guided module is incorporated to strengthen adaptation to new tasks and mitigate negative interference from outdated policies. Experimental results on incremental multiscenario navigation tasks indicate that the proposed method can, to some extent, alleviate forgetting, improve adaptability across scenarios, and enhance learning efficiency. Code repository: https://github.com/icode-qp/GCRL-ICMA.
PMID:
43019277
Bibliographic data and abstract were imported from PubMed on 13 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 3
- Comments 0