Search papers, labs, and topics across Lattice.
This paper introduces a bionic memory replay model that leverages the collaborative mechanisms of the hippocampus and prefrontal cortex to enhance decision-making in robots. By integrating meta-reinforcement learning with model-based policy optimization, the Memory Planning Replay (MPR) model enables robots to simulate future behaviors based on past experiences, significantly improving learning efficiency and adaptability in complex environments. Experimental validation demonstrates that MPR reduces navigation path lengths by 14% and achieves a success rate at least 10% higher than existing models, showcasing its potential for real-world applications in uncertain settings.
Robots using the MPR model can navigate complex environments with 14% shorter paths and significantly higher success rates, transforming how we approach autonomous decision-making.
The decision-making process in the human brain is not merely a function of existing memory, but an active cognitive process that intricately combines long-term empirical memory with forward-looking prediction. This process is supported by the cooperative interaction between the hippocampus and prefrontal cortex, which enables simulation of future behaviors. The brain simulates potential future behavioral paths and assesses their consequences, adjusting the depth of planning based on environmental structure and task requirements, thereby selecting the optimal decision within a limited time. Based on the hippocampal-prefrontal cooperative planning theory, a memory replay model driven by empirical value and environmental context is proposed. This model combines meta-reinforcement learning with model-based policy optimization techniques by establishing a closed-loop interaction between the hippocampal planning replay network and the prefrontal policy evaluation network. This bionic neural mechanism unifies dynamic programming and strategy optimization, and realizes the dynamic coupling of experience playback and prospective simulation in the process of brain incentive decision-making. Therefore, this framework enhances the agent’s learning ability and improves its learning efficiency in complex environments through “planning”. This paper verifies the proposed model through robot navigation experiments in both static and dynamic environments. The experimental results show that MPR reduces the length of navigation paths by 14%, and its success rate is at least 10% higher than that of other models. Moreover, in the early stage of training, the reward value of MPR was significantly higher than that of other models, and its learning efficiency was nearly twice as high as that of other models. These results verify the improved learning efficiency, convergence speed and adaptability of MPR, making it suitable for robot applications in uncertain and complex environments. Experimental results can be found at: https://github.com/LXDGRA/MPR Note to Practitioners—This work presents a brain-inspired decision-making framework for mobile robots that mimics the hippocampal-prefrontal interaction in the human brain to enhance planning and learning in unknown or dynamic environments. The proposed Memory Planning Replay (MPR) model integrates meta-reinforcement learning with model-based policy optimization, enabling robots to simulate future states and actions based on past experiences. This reduces reliance on large amounts of real-world interaction data and improves sample efficiency, convergence speed, and adaptability. Experimental results in both static and dynamic settings show that the MPR model outperforms existing methods like PERSAC, HER, and MBPO in terms of path optimality, success rate, and stability. Practitioners can leverage this model to improve autonomous navigation and task execution in complex environments where prior knowledge is limited or unavailable. Key implementation considerations include tuning the planning depth $L$ and balancing exploration versus exploitation through a dynamically weighted policy. The model is particularly suitable for applications requiring efficient and robust decision-making under uncertainty.
