Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Ling-Li Zeng has a total of 276 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Latent MeanFlow Policy Optimization for Offline Reinforcement Learning

    by Tenglong Liu, Xin Xu, Yixing Lan, Jiacheng Huang, Xinglong Zhang, Ling-Li Zeng

    Offline reinforcement learning (RL) aims to derive effective policies from fixed datasets without environment interaction. While generative models such as diffusion and flow matching improve policy expressiveness, they often suffer from high computational cost due to multi-step sampling and limited representational richness from uninformative Gaussian priors. To address these challenges, we propose Latent MeanFlow Policy Optimization (LaMPO), a generative policy framework that formulates offline RL as latent generative policy optimization. LaMPO learns a behavior-conditioned latent distribution, which provides an informative prior and reduces the modeling burden of the generative process. Conditioned on this prior, the MeanFlow policy is trained to predict the average action velocity, enabling high-fidelity one-step action generation that avoids the cost of iterative sampling. Guided by the MeanFlow policy, the target policy is then optimized under a MeanFlow-induced behavior constraint, achieving effective policy improvement. We further establish a theoretical performance lower bound of the target policy relative to the MeanFlow policy. Extensive evaluations across 71 tasks in OGBench, D4RL, and real-world humanoid manipulation demonstrate that LaMPO achieves superior performance with high efficiency, yielding a 19% average improvement over existing state-of-the-art methods and an 81% success rate in real-world robotic tasks.

  • Two-stage instance selection domain adaptation for diagnostic classification of neuropsychiatric disorders

    by Minghao Dai, Jianpo Su, Zhipeng Fan, Limin Peng, Dewen Hu, Ling-Li Zeng

  • Open Access

    Reinforcement learning in linear embedding space unlocks generalizable control across soft robot configurations

    by Xinglong Zhang, Cong Li, Hangjie Mo, Yue Jiang, Wenyu Cao, Xin Xu, Wei Jiang, Zhenshan Bing, Yihe Yang, Xiaojian Li, et al.

    Soft-bodied organisms exhibit prominent morphological adaptability, dynamically reconfiguring shape and stiffness to achieve versatile behaviors. Inspired by these systems, soft robots with diverse morphologies have emerged, yet a unified control framework that rapidly adapts across configurations remains elusive. Here, we introduce a generalizable control system that enables rapid cross-configuration adaptation via reinforcement learning in a shared linear Koopman embedding space. By encoding robot dynamics into this embedding space, our method decouples control policies from specific morphologies, allowing real-time, model-free policy adaptation without retraining from scratch. We validate our system across 33 distinct robot configurations. Our system achieves a 75 × reduction in transfer samples across configurations, while sustaining robust performance under high-speed motion, heavy payloads, and multiactuator faults, and achieving real-world skills previously unattainable in soft robotics. This work establishes an adaptable control framework for diverse soft robot configurations and may offer insights for generalizable control in complex physical systems.


  • Learning Task-Preferred Inference Routes for Gradient De-Conflict in Multi-Output DNNs

    by Yi Sun, Xiaochang Hu, Xin Xu, Jian Li, Yifei Shi, Ling-Li Zeng

    Multi-output deep neural networks (MONs) contain multiple output branches of various tasks, and these tasks typically share partial network filters, resulting in entangled inference routes between different tasks within the networks. Due to the divergent optimization objectives, the task gradients during training usually interfere with each other along the shared routes, which decreases the overall model performance. To address this issue, we propose a novel gradient de-conflict algorithm named DR-MGF (Dynamic Routes and Meta-weighted Gradient Fusion). Different from existing de-conflict methods, DR-MGF achieves gradient de-conflict in MONs by learning task-preferred inference routes. The proposed method is motivated by our experimental findings that the shared filters are not equally important for different tasks. By designing learnable task-specific importance variables, DR-MGF evaluates the importance of filters for different tasks. Through making the dominance of tasks over filters proportional to the task-specific importance of filters, DR-MGF can effectively reduce inter-task interference. These task-specific importance variables ultimately determine task-preferred inference routes at the end of training iterations. Extensive experimental results on CIFAR, ImageNet, and NYUv2 demonstrate that DR-MGF outperforms existing de-conflict methods. Furthermore, DR-MGF can be extended to general MONs without modifying the overall network structures.


  • Parallel Brain–Computer Interface: Collaborative Control and Perception Based on Multibrain Division of Labor

    by Zeqi Ye, Yang Yu, Yiyun Zhang, Zongtan Zhou, Dewen Hu, Ling-Li Zeng

    To date, brain-computer interfaces (BCIs) have adopted a serial working mechanism, causing performance bottlenecks due to time multiplication. In this study, we propose a parallel BCI (pBCI), which transforms the traditional sequential processing paradigm into multiple sub-paradigms based on division-labor mechanism, and then these sub-paradigms can be processed by multiple brains in a parallel and collaborative way. According to the proposed pBCI, we designed: 1) a Collaborative Speller paradigm where targets are encoded in a multi-dimensional space, with each brain focusing on a distinct dimension and collaborating on target localization; and 2) a Collaborative Perception paradigm with each brain contributing to a unique perspective of an autonomous vehicle and collaborating on perceiving dangers. We conducted both Collaborative P300 Speller experiments and Collaborative Perception experiments to validate the feasibility and practicality of the proposed pBCI. The results reveal that the proposed pBCI boosted BCI performance by multi-brain division-labor collaboration without increasing the workload for individuals, resulting in an information transfer rate improvement from 55.17bits/min to 146.35bits/min in Collaborative Speller, and a response time improvement from 1696ms to 1130ms in Collaborative Perception. In summary, the proposed pBCI provides a new and scalable BCI solution for more complex tasks and applications.

  • DMAE-EEG: A Pretraining Framework for EEG Spatiotemporal Representation Learning

    by Yifan Zhang, Yang Yu, Hao Li, Anqi Wu, Xin Chen, Jinfang Liu, Ling-Li Zeng, Dewen Hu

    Electroencephalography (EEG) plays a crucial role in neuroscience research and clinical practice, but it remains limited by nonuniform data, noise, and difficulty in labeling. To address these challenges, we develop a pretraining framework named DMAE-EEG, a denoising masked autoencoder for mining generalizable spatiotemporal representation from massive unlabeled EEG. First, we propose a novel brain region topological heterogeneity (BRTH) division method to partition the nonuniform data into fixed patches based on neuroscientific priors. Second, we design a denoised pseudo-label generator (DPLG), which utilizes a denoising reconstruction pretext task to enable the learning of generalizable representations from massive unlabeled EEG, suppressing the influence of noise and artifacts. Furthermore, we utilize an asymmetric autoencoder with self-attention as the backbone in the proposed DMAE-EEG, which captures long-range spatiotemporal dependencies and interactions from unlabeled EEG data across 14 public datasets. The proposed DMAE-EEG is validated on both generative (signal quality enhancement) and discriminative tasks (motion intention recognition). In the quality enhancement, DMAE-EEG outperforms existing statistical methods with normalized mean squared error (nMSE) reduction of 27.78%–50.00% under corruption levels of 25%, 50%, and 75%, respectively. In motion intention recognition, DMAE-EEG achieves a relative improvement of 2.71%–6.14% in intrasession classification balanced accuracy across 2–6 class motor execution and imagery tasks, outperforming state-of-the-art methods. Overall, the results suggest that the pretraining framework DMAE-EEG can capture generalizable spatiotemporal representations from massive unlabeled EEG and enhance the knowledge transferability across sessions, subjects, and tasks in various downstream scenarios, advancing EEG-aided diagnosis and brain–computer communication and control, and other clinical practice.


  • Feature and Semantic Matching Multisource Domain Adaptation for Diagnostic Classification of Neuropsychiatric Disorders

    by Minghao Dai, Jianpo Su, Zhipeng Fan, Chenyu Wang, Limin Peng, Dewen Hu, Ling-Li Zeng

    Multi-site resting-state functional magnetic resonance imaging (rs-fMRI) data have been increasingly utilized for diagnostic classification of neuropsychiatric disorders such as schizophrenia (SCZ) and major depressive disorder (MDD). However, the cross-site generalization ability of deep networks is limited due to the significant inter-site data heterogeneity caused by different MRI scanners or scanning protocols. To address this issue, we propose a feature and semantic matching multi-source domain adaptation method (FSM-MSDA) to learn site-invariant disorder-related feature representations. In FSM-MSDA, we adopt separate feature extractors for multiple source domains, and propose an accurate feature matching module to align the category-level feature distributions across multiple source domains and target domain. In addition, we also propose a semantic feature alignment module to eliminate the distribution discrepancy in high-level semantic features extracted from target samples by different source classifiers. Extensive experiments based on multi-site fMRI data of SCZ and MDD show the superiority and robustness of FSM-MSDA compared to state-of- the-art methods. Besides, FSM-MSDA achieves the average accuracy of 80.8% in the classification of SCZ, meeting the clinically diagnostic accuracy threshold of 80%. Shared discriminative brain regions including the middle temporal gyrus and the cerebellum regions are identified in the diagnostic classification of SCZ and MDD.


  • Open Access

    Universal slip detection of robotic hand with tactile sensing

    by Chuangri Zhao, Yang Yu, Zeqi Ye, Ziyang Tian, Yifan Zhang, Ling-Li Zeng

    Slip detection is to recognize whether an object remains stable during grasping, which can significantly enhance manipulation dexterity. In this study, we explore slip detection for five-finger robotic hands being capable of performing various grasp types, and detect slippage across all five fingers as a whole rather than concentrating on individual fingertips. First, we constructed a dataset collected during the grasping of common objects from daily life across six grasp types, comprising more than 200 k data points. Second, according to the principle of deep double descent, we designed a lightweight universal slip detection convolutional network for different grasp types (USDConvNet-DG) to classify grasp states (no-touch, slipping, and stable grasp). By combining frequency with time domain features, the network achieves a computation time of only 1.26 ms and an average accuracy of over 97% on both the validation and test datasets, demonstrating strong generalization capabilities. Furthermore, we validated the proposed USDConvNet-DG in real-time grasp force adjustment in real-world scenarios, showing that it can effectively improve the stability and reliability of robotic manipulation.