Frontiers in Neurorobotics最新文献

The development of air traffic control (ATC) automation has been constrained by the scarcity and low quality of communication data, particularly in low-altitude complex airspace, where non-standardized instructions frequently hinder training efficiency and operational safety. This paper proposes the BART-Reinforcement Learning (BRL) model, a deep reinforcement learning model based on the BART pre-trained language model, optimized through transfer learning and reinforcement learning techniques. The model was evaluated on multiple ATC datasets, including training flight data, civil aviation operational data, and standardized datasets generated from Radiotelephony Communications for Air Traffic Services. Evaluation metrics included ROUGE and semantic intent-based indicators, with comparative analysis against several baseline models. Experimental results demonstrate that BRL achieves a 10.5% improvement in overall accuracy on the training dataset with the highest degree of non-standardization, significantly outperforming the baseline models. Furthermore, comprehensive evaluations validate the model's effectiveness in standardizing various types of instructions. The findings suggest that reinforcement learning-based approaches have the potential to significantly enhance ATC automation, reducing communication inconsistencies, and improving training efficiency and operational safety. Future research may further optimize standardization by incorporating additional contextual factors into the model.

Introduction: Foundation pit deformation prediction is a critical aspect of underground engineering safety assessment, influencing construction quality and personnel safety. However, due to complex geological conditions and numerous environmental interference factors, traditional prediction methods struggle to achieve precise modeling. Conventional approaches, including numerical simulations, empirical formulas, and machine learning models, suffer from limitations such as high computational costs, poor generalization, or excessive dependence on specific data distributions. Recently, deep learning models, particularly cross-modal architectures, have demonstrated great potential in engineering applications. However, effectively integrating multi-modal data for improved prediction accuracy remains a significant challenge.

Methods: This study proposes a Multi-Scale Contrastive Language-Image Pretraining (CLP) framework, ModuCLIP, designed for foundation pit deformation prediction in multi-modal robotic systems. The framework leverages a self-supervised contrastive learning mechanism to integrate multi-source information, including images, textual descriptions, and sensor data, while employing a multi-scale feature learning approach to enhance adaptability to complex conditions. Experiments conducted on multiple foundation pit engineering datasets demonstrate that ModuCLIP outperforms existing methods in terms of prediction accuracy, generalization, and robustness.

Results and discussion: The findings suggest that this framework provides an efficient and precise solution for foundation pit deformation prediction while offering new insights into multi-modal robotic perception and engineering monitoring applications.

[This corrects the article DOI: 10.3389/fnbot.2025.1482281.].

This study addresses the distributed optimization problem with time-varying objective functions and time-varying constraints in a multi-agent system (MAS). To tackle the distributed time-varying constrained optimization (DTVCO) problem, each agent in the MAS communicates with its neighbors while relying solely on local information, such as its own objective function and constraints, to compute the optimal solution. We propose a novel penalty-based zeroing neural network (PB-ZNN) to solve the continuous-time DTVCO (CTDTVCO) problem. The PB-ZNN model incorporates two penalty functions: The first penalizes agents for deviating from the states of their neighbors, driving all agents to reach a consensus, and the second penalizes agents for falling outside the feasible range, ensuring that the solutions of all agents remain within the constraints. The PB-ZNN model solves the CTDTVCO problem in a semi-centralized manner, where information exchange between agents is distributed, but computation is centralized. Building on the semi-centralized PB-ZNN model, we adopt the Euler formula to develop a distributed PB-ZNN (DPB-ZNN) algorithm for solving the discrete-time DTVCO (DTDTVCO) problem in a fully distributed manner. We present and prove the convergence theorems of the proposed PB-ZNN model and DPB-ZNN algorithm. The efficacy and accuracy of the DPB-ZNN algorithm are illustrated through numerical examples, including a simulation experiment applying the algorithm to the cooperative control of redundant manipulators.

Event-based cameras are bio-inspired vision sensors that mimic the sparse and asynchronous activation of the animal retina, offering advantages such as low latency and low computational load in various robotic applications. However, despite their inherent sparsity, most existing visual processing algorithms are optimized for conventional standard cameras and dense images captured from them, resulting in computational redundancy and high latency when applied to event-based cameras. To address this gap, we propose a sparse convolution operator tailored for event-based cameras. By selectively skipping invalid sub-convolutions and efficiently reorganizing valid computations, our operator reduces computational workload by nearly 90% and achieves almost 2× acceleration in processing speed, while maintaining the same accuracy as dense convolution operators. This innovation unlocks the potential of event-based cameras in applications such as autonomous navigation, real-time object tracking, and industrial inspection, enabling low-latency and high-efficiency perception in resource-constrained robotic systems.

Objective: To address the limitations of traditional methods in human pose recognition, such as occlusions, lighting variations, and motion continuity, particularly in complex dynamic environments for seamless human-robot interaction.

Method: We propose PoseRL-Net, a deep learning-based pose recognition model that enhances accuracy and robustness in human pose estimation. PoseRL-Net integrates multiple components, including a Spatial-Temporal Graph Convolutional Network (STGCN), attention mechanism, Gated Recurrent Unit (GRU) module, pose refinement, and symmetry constraints. The STGCN extracts spatial and temporal features, the attention mechanism focuses on key pose features, the GRU ensures temporal consistency, and the refinement and symmetry constraints improve structural plausibility and stability.

Results: Extensive experiments conducted on the Human3.6M and MPI-INF-3DHP datasets demonstrate that PoseRL-Net outperforms existing state-of-the-art models on key metrics such as MPIPE and P-MPIPE, showcasing superior performance across various pose recognition tasks.

Conclusion: PoseRL-Net not only improves pose estimation accuracy but also provides crucial support for intelligent decision-making and motion planning in robots operating in dynamic and complex scenarios, offering significant practical value for collaborative robotics.

High-speed walking is fundamental for humanoid robots to quickly reach the work site in emergency scenarios. According to biological studies, the coordinated motion of the arms and waist can significantly enhance walking speed and stability in humans. However, existing humanoid robot walking control frameworks predominantly focus on leg control, often overlooking the utilization of upper body joints. In this paper, a novel walking control framework combining the improved footstep planner and the whole-body coordination controller is proposed, aiming to improve the humanoid robot's tracking accuracy of desired speeds and its dynamic walking capability. First, we analyze the issues in traditional footstep planners based on Linear Inverted Pendulum and Model Predictive Control (LIP-MPC). By reconstructing the footstep optimization problem during walking using the Center-of-Mass (CoM) position, we propose an improved footstep planner to enhance the control accuracy of the desired walking speed in humanoid robots. Next, based on biological research, we define a coordinated control strategy for the arms and waist during walking. Specifically, the waist increases the robot's step length, while the arms counteract disturbance momentum and maintain balance. Based on the aforementioned strategy, we design a whole-body coordination controller for the humanoid robot. This controller adopts a novel hierarchical design approach, in which the dynamics and motion controllers for the upper and lower body are modeled and managed separately. This helps avoid the issue of poor control performance caused by multi-task coupling in traditional whole-body controllers. Finally, we integrate these controllers into a novel walking control framework and validate it on the simulation prototype of the humanoid robot Dexbot. Simulation results show that the proposed framework significantly enhances the maximum walking capability of the humanoid robot, demonstrating its feasibility and effectiveness.

Autonomous driving technology has garnered significant attention due to its potential to revolutionize transportation through advanced robotic systems. Despite optimistic projections for commercial deployment, the development of sophisticated autonomous driving systems remains largely experimental, with the effectiveness of neurorobotics-based decision-making and planning algorithms being crucial for success. This paper delivers a comprehensive review of decision-making and planning algorithms in autonomous driving, covering both knowledge-driven and data-driven approaches. For knowledge-driven methods, this paper explores independent decision-making systems, including rule based, state transition based, game-theory based methods and independent planing systems including search based, sampling based, and optimization based methods. For data-driven methods, it provides a detailed analysis of machine learning paradigms such as imitation learning, reinforcement learning, and inverse reinforcement learning. Furthermore, the paper discusses hybrid models that amalgamate the strengths of both data-driven and knowledge-driven approaches, offering insights into their implementation and challenges. By evaluating experimental platforms, this paper guides the selection of appropriate testing and validation strategies. Through comparative analysis, this paper elucidates the advantages and disadvantages of each method, facilitating the design of more robust autonomous driving systems. Finally, this paper addresses current challenges and offers a perspective on future developments in this rapidly evolving field.

Aiming at the problems of slow network convergence, poor reward convergence stability, and low path planning efficiency of traditional deep reinforcement learning algorithms, this paper proposes a BiLSTM-D3QN (Bidirectional Long and Short-Term Memory Dueling Double Deep Q-Network) path planning algorithm based on the DDQN (Double Deep Q-Network) decision model. Firstly, a Bidirectional Long Short-Term Memory network (BiLSTM) is introduced to make the network have memory, increase the stability of decision making and make the reward converge more stably; secondly, Dueling Network is introduced to further solve the problem of overestimating the Q-value of the neural network, which makes the network able to be updated quickly; Adaptive reprioritization based on the frequency penalty function is proposed. Experience Playback, which extracts important and fresh data from the experience pool to accelerate the convergence of the neural network; finally, an adaptive action selection mechanism is introduced to further optimize the action exploration. Simulation experiments show that the BiLSTM-D3QN path planning algorithm outperforms the traditional Deep Reinforcement Learning algorithm in terms of network convergence speed, planning efficiency, stability of reward convergence, and success rate in simple environments; in complex environments, the path length of BiLSTM-D3QN is 20 m shorter than that of the improved ERDDQN (Experience Replay Double Deep Q-Network) algorithm, the number of turning points is 7 fewer, the planning time is 0.54 s shorter, and the success rate is 10.4% higher. The superiority of the BiLSTM-D3QN algorithm in terms of network convergence speed and path planning performance is demonstrated.

Human skeleton-based action recognition is an important task in the field of computer vision. In recent years, masked autoencoder (MAE) has been used in various fields due to its powerful self-supervised learning ability and has achieved good results in masked data reconstruction tasks. However, in visual classification tasks such as action recognition, the limited ability of the encoder to learn features in the autoencoder structure results in poor classification performance. We propose to enhance the encoder's feature extraction ability in classification tasks by leveraging the latent space of variational autoencoder (VAE) and further replace it with the latent space of vector quantized variational autoencoder (VQVAE). The constructed models are called SkeletonMVAE and SkeletonMVQVAE, respectively. In SkeletonMVAE, we constrain the latent variables to represent features in the form of distributions. In SkeletonMVQVAE, we discretize the latent variables. These help the encoder learn deeper data structures and more discriminative and generalized feature representations. The experiment results on the NTU-60 and NTU-120 datasets demonstrate that our proposed method can effectively improve the classification accuracy of the encoder in classification tasks and its generalization ability in the case of few labeled data. SkeletonMVAE exhibits stronger classification ability, while SkeletonMVQVAE exhibits stronger generalization in situations with fewer labeled data.