Extensive studies have shown that many animals’ capability of forming spatial representations for self-localization, path planning, and navigation relies on the functionalities of place and head-direction (HD) cells in the hippocampus. Although there are numerous hippocampal modeling approaches, only a few span the wide functionalities ranging from processing raw sensory signals to planning and action generation. This paper presents a vision-based navigation system that involves generating place and HD cells through learning from visual images, building topological maps based on learned cell representations and performing navigation using hierarchical reinforcement learning. First, place and HD cells are trained from sequences of visual stimuli in an unsupervised learning fashion. A modified Slow Feature Analysis (SFA) algorithm is proposed to learn different cell types in an intentional way by restricting their learning to separate phases of the spatial exploration. Then, to extract the encoded metric information from these unsupervised learning representations, a self-organized learning algorithm is adopted to learn over the emerged cell activities and to generate topological maps that reveal the topology of the environment and information about a robot’s head direction, respectively. This enables the robot to perform self-localization and orientation detection based on the generated maps. Finally, goal-directed navigation is performed using reinforcement learning in continuous state spaces which are represented by the population activities of place cells. In particular, considering that the topological map provides a natural hierarchical representation of the environment, hierarchical reinforcement learning (HRL) is used to exploit this hierarchy to accelerate learning. The HRL works on different spatial scales, where a high-level policy learns to select subgoals and a low-level policy learns over primitive actions to specialize on the selected subgoals. Experimental results demonstrate that our system is able to navigate a robot to the desired position effectively, and the HRL shows a much better learning performance than the standard RL in solving our navigation tasks.

Synthetic data has become essential in machine learning and data science, addressing real-world data limitations such as scarcity, privacy, and cost. While existing generative models are effective in synthesizing various sensor data, they struggle with performance and generalization. This letter introduces a large language model (LLM)-driven framework that leverages LLMs and domain-specific generative models (DGMs) for general sensor data synthesis. Specifically, our method employs LLMs as the core to analyze data generation tasks, decompose complex tasks into manageable subtasks, and delegate each to the most suitable DGM, thereby automatically constructing customized data generation pipelines. Meanwhile, the integration of reinforcement learning (RL) is promising to enhance the framework's ability to optimally utilize DGMs, resulting in data generation with superior quality and control flexibility. Experimental results demonstrate the effectiveness of LLMs in understanding diverse tasks and in facilitating general sensor data synthesis through collaborative interactions with diverse DGMs.

Digital twins serve as vital tools for monitoring and simulating real-world systems, yet ensuring their accuracy and adaptability in dynamic scenarios remains a challenge. In this paper, we introduce FlexiTwin, a digital twin updating assistance platform seamlessly integrated with AdaSor, a lightweight adaptive data selector. FlexiTwin automates the construction of incremental learning datasets for updating digital twins within specified time constraints, ensuring their adaptation to new scenarios while preserving historical knowledge. Through simulations focusing on UAV energy management, we show that FlexiTwin substantially enhances the adaptability of digital twins to new scenarios while effectively preserving their accuracy in historical scenarios.

Existing video-based human pose estimation methods commonly rely on large networks to localize body joints across all frames, achieving remarkable accuracy but imposing high memory and computational demands that limit their applications on resource-constrained devices. Moreover, most models lack the capability to accommodate dynamic changes in available resources, which can negatively impact the performance of parallel tasks. To address these issues, this paper proposes a novel yet effective framework for efficient and adaptive human pose estimation on resource-constrained devices. Specifically, the proposed approach adopts the knowledge distillation (KD) strategy to train a light-weight pose estimator network, which is capable of executing rapidly with low computational cost. To further increase the overall efficiency, it exploits the temporal coherence between successive video frames and explicitly propagates body joints from previous frames rather than naively extracting them using a pose estimator. Furthermore, a prediction-based mechanism is adopted to facilitate adaptive key-frame selection, dynamically determining the optimal number of keyframes, thus enhancing the overall efficiency and adaptability. Experiments on Penn Action, Sub-JHMDB, and real-world systems demonstrate that the proposed method achieves comparative accuracy, superior efficiency, and robust flexibility in dynamic scenarios.

Federated learning (FL) has emerged as a pivotal approach for edge-based distributed machine learning, yet it faces significant challenges due to the constrained capacities and heterogeneity of edge devices, including non-IID data distribution, communication constraints, and learning inefficiencies. Furthermore, a one-fits-all global model often fails to perform optimally across diverse participating devices. In this article, we present NestFL, an efficient FL framework for edge computing that can jointly improve the training efficiency and achieve personalization. Specifically, NestFL innovates by incorporating distributed model pruning, creating a hierarchy of structured-sparse subnetworks tailored to the unique resource profiles of client devices. These subnetworks are integrated into a nested global model, ensuring parameter sharing without increasing the parameter space, thereby significantly reducing computational and communication burdens. Meanwhile, it implements a cross-training mechanism, allowing clients to train on a broader dataset and maintain consistent decision boundaries. Furthermore, a weighted aggregation mechanism is designed to improve training performance and maximally preserve personalization. Experimental results in different applications demonstrate the superiority of NestFL over the baseline approaches in terms of model accuracy, convergence speed, and personalization preservation.

This article presents AdaHPE, an innovative and efficient framework for human pose estimation (HPE) designed specifically for edge computing devices with constrained and fluctuating resources. AdaHPE redefines the conventional HPE workflow by converting the resource-demanding pose regression into a sequence of computationally feasible pose propagation tasks. The framework incorporates a memory-augmented long short-term memory (LSTM) network with a global memory repository, allowing AdaHPE to adaptively choose keyframes based on real-time data and the device’s resource status, thereby optimizing the tradeoff between accuracy and computational efficiency. A reinforcement learning component is further integrated to intelligently adjust the ratio of keyframes used, enhancing the framework’s adaptability. Utilizing policy gradient algorithms, AdaHPE is optimized to maximize a reward function that encourages both accurate and resource-efficient pose estimations, while respecting a given keyframe constraint. Extensive experiments on benchmarks, including Penn Action, Sub-JHMDB, NTU RGB+D 120, and real-world datasets demonstrate that AdaHPE can significantly reduce computational overhead compared to per-frame HPE models while preserving high accuracy and robustness under varying resource limitations. Moreover, the seamless compatibility of our approach with various off-the-shelf HPE models highlights its versatility and potential for broad applications.