Weiyong Si

Robot-assisted ultrasound scanning has been proven to be an effective solution for medical examination; however, most of the existing work is fully teleoperation or semiautonomous due to dynamic tasks and uncertain environments. In this article, we propose a novel robot learning system for autonomous ultrasound scanning that integrates dynamic target localization with motion generation based on learning from demonstration (LfD). First, a new LfD model, Gaussian process dynamic movement primitives (GPDMP), is introduced to encode human motion skills and generate motion trajectories online. The proposed GPDMP framework supports multiple demonstrations and guarantees convergence toward the target even in the presence of external disturbances, such as subject movement during ultrasound scanning. To further enhance generalization to moving subjects through robust target localization, a vision-based localization method is developed to identify and localize the neck region. Neck positioning across individuals with different body shapes is addressed by an improved AlphaPose network, while interference from other moving subjects in the scene is mitigated using an enhanced DeepSORT-based tracking algorithm. The proposed system is capable of automatically identifying and localizing the human neck, generating approach trajectories for the ultrasound probe, and achieving real-time interaction. A robot-assisted neck ultrasound scanning system is implemented to experimentally demonstrate the effectiveness of the proposed framework.

Drawing inspiration from the brain's neurocognitive mechanisms of information chunking and topographic mapping, adaptive decision-making requires neural-grounded architectures that are interpretable and resilient to uncertainty. In this paper, we propose a novel boosted fuzzy manifold granule hypersurface classifier (BFMGHC). The algorithm performs classification at the “information granule" level, realizing an intelligent modeling method that is closer to human cognition, more interpretable, and more accommodating of uncertainty. The classifier mainly consists of three main parts: (1) A manifold-based measurement method for samples that preserves local topological structure is designed, echoing the topographic representations in neural dynamics. Based on this, a global optimization clustering algorithm is proposed and integrated with the Dask framework to achieve scalable hierarchical parallel granulation from raw inputs to high-level semantic granules. (2) In the fuzzy manifold granule space, a measurement method and a hypersurface classifier are constructed, utilizing a particle swarm optimization method for parameter solving. (3) To improve interpretability, weights are assigned to different granules and base classifiers, resembling bio-inspired neuromodulation to ensure stable behavior. The proposed BFMGHC was verified on three financial risk assessment datasets in the UCI Machine Learning Repository (Default of Credit Card Clients, Bank Marketing, and German Credit Data) and achieved superior performance.

The convergence of foundation models, including Large Language Models (LLMs), Vision Language Action (VLA) models, with robotics and computational intelligence represents one of the most significant developments in modern AI research. Traditional robotic systems, while capable in structured environments, have long struggled with the safety, intuitive interaction, and adaptive reasoning required for real-world deployment. By integrating the natural language understanding and contextual reasoning capabilities of LLMs into robotic frameworks, researchers are now unlocking new possibilities spanning voice-driven manipulation, autonomous navigation, and human-robot interaction. This Research Topic brings together five original research articles that explore the growing role 23 of LLMs in robotic and intelligent systems. The contributions cover a broad range of applications, 24 including supermarket assistance robots, bimanual teleoperation, home-care navigation, educational 25 robots, and automated machine learning, but share a common focus: using the language and reasoning 26 capabilities of LLMs to bridge the gap between human intention and robot action. These works offer new 27 methodologies, system architectures, and empirical findings that advance our understanding of how LLMs 28 and computational intelligence can be effectively used in advanced robotic systems. achieving the lowest error rates, and language found to be a more significant source of variation than gender.Building on this, the authors propose a hierarchical multi-LLM conversational agent where a fine-tuned 34 DistilBERT classifier routes user queries to specialised GPT-3.5 Turbo models, grounded to a product 35 database using Retrieval-Augmented Fine-Tuning to reduce hallucinations. When compared against a 36 GPT-4 Turbo benchmark, the proposed system performs better across all evaluated criteria. The work also 37 demonstrates integration with a mobile robot for autonomous shelf navigation, suggesting that coordinated 38 ensembles of smaller, task-specific models can be a practical alternative to relying on a single large model.

Autonomous dynamic systems (ADS)-based encoding of human motion skills has been widely demonstrated to exhibit high transfer efficiency in goal-directed tasks; however, significant gaps remain in the study of skill transfer for periodic motions, particularly complex periodic motions. In this letter, we propose a novel method for directly learning complex periodic movements using Lyapunov functions (LF), which is fundamentally different from all existing methods that learn periodic movements through limit cycle mapping. First, we introduce a polar coordinate transformation to decouple trajectory and angle modulation, enabling the radially increasing Lyapunov energy function to acquire non-radial expansion capabilities. Then, based on this, we learn a data-driven Lyapunov energy function by solving a dual objective optimization problem that combines geometric alignment and orbit parallelism constraints, aligning one of its horizontal surfaces with a periodic human demonstration trajectory. Finally, ADS is learned by sequentially solving LF-related constrained optimization problems. By designing appropriate constraint functions, we can ensure that the trajectories generated by ADS converge to an LF-level surface whose shape resembles that of periodic human demonstration trajectories. Both simulations and real robot experiments validate the effectiveness of the proposed method.

Reliable pose estimation of non-cooperative spacecraft is a key technology for on-orbit servicing and active space debris removal missions. Currently, deep learning has become the mainstream method for spacecraft pose estimation. However, existing methods suffer from problems such as excessive parameters, high computational complexity, and relatively low efficiency in feature utilization. This letter proposes SDPENetv2 to address these issues. We represent the pose of the target spacecraft using discrete pose weights. By introducing an additional constraint term into the loss function, the network can learn a more appropriate weight distribution in the later stages of training, improving the accuracy of pose estimation. In addition, we propose the learnable token head, which possesses global attention and can make more comprehensive use of the features extracted by the convolutional neural network. Experiments on the SPEED dataset demonstrate that the position and attitude estimation error of SDPENetv2 are reduced to 0.105 m and 1.145$^{\circ }$, respectively. Compared with other works, these errors are reduced by 20.45% –86.59% and 32.65% –91.80%, respectively. Additionally, SDPENetv2 has only 5.6 M parameters and a computational complexity of merely 1.542 GMACs.

Medical ultrasound examination is a challenging dexterous manipulation task for robots. Even for experienced sonographers, it involves real-time decision-making, motion control, and force regulation based on ultrasound images and patient feedback. In this article, we propose a unified framework for robot-assisted medical examination, specifically for the initial registration in artery scanning, by leveraging deep multimodal imitation learning, compliant control, and trajectory optimization. To process multimodal inputs during the initial registration phase, we investigate a deep imitation learning model that fuses RGB and ultrasound images, contact force, and proprioceptive data. The deep imitation learning model predicts the desired motion and contact force. We design a compliant controller in Cartesian space to track the desired trajectory and force. To smooth the trajectory and ensure safety, we employ a trajectory optimization planner between the deep imitation learning module and the low-level compliant controller. The generalization capability of the deep multimodal imitation learning module, control performance, and the quality of the acquired ultrasound images on both the Phantom and human subjects were evaluated. Experimental results show that the proposed approach significantly improves the success rate of autonomous ultrasound scanning from 75% to 90%, while also reducing the completion time.