The advent of cellular 5G technologies introduces positioning as a new service, poised to enhance robotics applications through mobile communications in both urban and rural settings. Field robots in outdoor environments, are faced with unreliable GNSS coverage in urban/rural canyons and obstructions. In featureless terrains and under poor illumination, optical sensing may suffer from significant drift, impairing robot localization. We propose NR5G-SAM-LC, a radio SLAM framework which is signal-independent, capable of utilizing 5GNR, WiFi, or UWB signals, functioning as a standalone or complementary pose information source to these problems. The framework’s efficacy is demonstrated using an UWB testbed, chosen for its similar signal structure and propagation characteristics to 5GNR systems in the FR1 bands and replicate irregular cellular station geometry placement in urban scenarios. The novelty of this work lies in the exploitation of LoS, ambiguous LoS, NLoS conditions, along with station availability and signal strength vectors to improve pose estimation. A multi-link Channel State Information (CSI) interpolation algorithm and radio loop closure methodology is developed that models Received Signal Strength Indicator (RSSI) measurements as radio point clouds, enabling the construction of Radio Environmental Maps (REM). Radio loop closures (LC) are explored as a potential localization aid for radio factor graph SLAM systems. Field experiments using a UGV indicate that REM can play a key role in supporting robotic systems by complementing pose estimation robustness with the application of a spatial transform, and radio loop closures can be used to detect previously visited locations based on similar signal signatures with varying degree of accuracy. Finally, the system is evaluated by comparing the individual block elements, UEB, BEB-REM, BEB-LC and FEB with onboard GNSS RTK and a LiDAR SLAM approach, highlighting its potential for enhancing the robustness of robotic autonomy in challenging real-world scenarios. Results indicate that REM in SLAM combined with radio loop closure capabilities of multi link signatures, can reduce the localization RMSE by 11%, compared to range only localization and radio loop closures can occur at previously visited locations reducing long term navigation errors and temporary drift in multimodal SLAM systems.
Snake robots generate locomotion through many simultaneous ground contacts, making them a challenging platform for model-based control. Contact-implicit trajectory optimization has shown success for legged and manipulation systems using rigid contact models, but its application to snake robots has been limited due to the complexity of distributed, contact-rich interactions. In this work, we propose a modeling approach that extends rigid contact formulations to snake robot locomotion and demonstrate that such models can also be adapted to deformable terrain. We evaluate this framework using two complementary simulation tools—Simscape Multibody for rigid ground and Chrono’s Soil Contact Model for compliant terrain, and validate the predictions through experiments on rigid ground and sand. Results from both simulations and hardware experiments show that across multiple gaits and gait frequencies, both the rigid and compliant ground models capture the dominant locomotion signatures, direction of travel, and head motion trajectories observed in experiments.
The automation of construction processes using robotic systems promises considerable increases in efficiency. However, a key challenge lies in the path planning of the tool center point (TCP), taking into account the complex environments on construction sites. Conventional methods for inverse kinematics (IKs) and reachability analysis often reach their limits in terms of flexibility and are less suitable for incorporation into gradient-based optimization. Especially when coordinating multiple robots, simultaneous optimization of TCP poses and assignment is crucial to ensure effective execution. This paper investigates neural networks (NNs) to determine the IKs, compares different network architectures and the effect of positional encoding for manipulators with multiple solutions of the IK. Additionally, an NN for predicting the kinematic reachability is presented. For both NNs, it is shown that encoding the positional values is particularly advantageous for robots with a large workspace and tasks that involve little or no redundancy. Based on the NNs, an augmented Lagrangian optimization problem for planning TCP poses for component transportation is designed, which jointly optimizes path poses and latent IK variables within a unified framework. The optimization takes into account collisions, kinematic reachability, number of handovers, joint configuration changes and path smoothness. The method is examined on three simulative test cases using a manipulator arm from Jekko and the UR10e from Universal Robots. These include simultaneous optimization of multiple paths, obstacle avoidance and assignment of path sections to a specific robot. Compared to sampling-based planners such as IRRT⋆ and BIT⋆, the proposed method achieves lower computation times in the majority of evaluated scenarios while simultaneously producing superior path quality in terms of path length, orientation consistency, and joint configuration changes.
The growing emphasis on sustainable waste management has increased the need for intelligent robotic systems capable of autonomously identifying and sorting recyclable materials. This paper presents a hybrid control framework for an autonomous mobile manipulator specifically designed for metal recycling applications. The proposed system integrates a deep learning–based perception module with a robust hybrid control strategy to enable efficient and precise handling of target objects in unstructured environments. A convolutional neural network (You Only Look Once, version 8 nano) is employed for real-time classification and localization of metal components, such as aluminum, allowing accurate target recognition under variable lighting and cluttered backgrounds. The control framework combines model-based trajectory planning with complementary fuzzy controllers and adaptive feedback mechanisms to ensure stable and responsive manipulation during object grasping and placement. The proposed architecture enhances adaptability to uncertainties in object position, shape, and orientation, which are common in real-world recycling scenarios. The system was validated through extensive experiments conducted in a simulated recycling environment and tested under real-world conditions. Performance metrics such as classification accuracy, grasping success rate, sorting efficiency, and task completion time were used to evaluate system effectiveness. Results demonstrate a significant improvement in autonomous sorting performance, with enhanced operational reliability and reduced reliance on manual labor. This research highlights the potential of intelligent robotic systems to transform industrial recycling by improving efficiency, reducing occupational hazards, and supporting sustainable development goals. The proposed framework offers a scalable solution for deploying autonomous robots in complex and dynamic waste management environments.
This paper presents an adaptive geometric proportional-derivative (PD) attitude controller for quadrotors on SO(3) that estimates the six independent inertia parameters online through a linear-in-parameters regressor. The estimator combines tracking-error-driven gradient descent with a concurrent-learning term that reuses recorded input–output data through an acceleration-free filtered regressor. This scheme guarantees exponential parameter convergence at a tunable rate under a finite-excitation condition, a weaker condition than persistent excitation, without external dither. A single Lyapunov function establishes boundedness and convergence under bounded disturbances, while robustness is enforced by σ-modification and projection onto the symmetric positive-definite cone. Numerical validation against a non-adaptive baseline in which the estimate remains frozen at a mismatched initial value shows that online adaptation is essential, in particular reducing the steady parameter error from over 70% to a few percent under sustained excitation. In addition, this scheme decreases the RMS attitude error by factors of approximately two to six. Concurrent learning further accelerates convergence by roughly a factor of four relative to a gradient-only law, converges from a finite excitation burst where the gradient law stalls, and tracks stepwise payload changes through age-based history-stack management. An operation count, a memory-footprint estimate, and a real-time margin analysis confirm that these benefits are obtained at a cost compatible with embedded flight-control hardware.
The COVID-19 pandemic caused millions of deaths and lasting societal impacts, showing that traditional swab collection methods place heavy burdens and infection risks on healthcare workers (HCWs). Recently developed sampling robots offer alternatives, but the possibility of contact-related infection remains. This study presents a robotic nasopharyngeal (NP) specimen collection system designed for automated NP sampling with limited HCW involvement. The system integrates a booth-structured enclosure, robotic manipulator, collaborative modules, and automation workflow. The robot controller combines camera-based visual servoing for initial swab alignment with compliant feedback control for safe and accurate insertion. Experiments confirmed precise alignment, safe force regulation, and completion of a sampling cycle within acceptable time and contact force. A survey of healthcare professionals reported high satisfaction with the contactless environment, disinfection system, and workload reduction. The proposed system demonstrates the feasibility of highly automated NP sampling for minimizing direct HCW exposure to infectious agents during large-scale specimen collection.
While Vision-Language Models (VLMs) have advanced embodied navigation, deploying lightweight models on resource-constrained platforms remains challenging due to perception hallucinations and weak command compliance. To address these issues, we propose the Semantically-Gated Visual Servo (SGVS) framework, a dual-system cognitive architecture tailored for lightweight VLMs. The framework integrates a Visual Object Interception (VOI) mechanism as a reflexive system to trigger rapid maneuvers upon target recognition, enhancing execution efficiency. Concurrently, a Closed-Loop Semantic Error Correction (CL-SEC) mechanism acts as a reflective system. By combining multi-prompt verification with dynamic negative sample mining, CL-SEC translates perceptual errors into real-time episodic memory and employs action pruning to block hallucination-driven failures at the source. Experimental results on the HM3D dataset demonstrate that SGVS improves the navigation success rate of the Qwen2.5-VL model from 45.7% to 48.0% while reducing the dangerous false positive rate by 30.4%. This study indicates that system-level architectural interventions can partially compensate for the perceptual limitations of lightweight foundational models.
Real-time replanning for dynamic obstacle avoidance is a critical issue in mobile robotics. While probabilistic roadmaps (PRMs) are effective for static obstacle navigation, they struggle to adapt efficiently to dynamic scenarios. We present the Deform-PRM technique, a real-time dynamic obstacle-avoidance approach that uses a local path-reshaping strategy within the framework of probabilistic roadmap-based planning. In this approach, a repulsive pseudo-force derived primarily from static and dynamic obstacles in the environment is applied to the sampled milestones that collide with the predicted collision state of dynamic obstacles. The collision-state prediction with dynamic obstacles is determined using a constant-velocity model of the dynamic obstacles and a forward-looking collision prediction technique. In addition to ensuring obstacle avoidance, our pseudo-force function promotes socially compliant navigation by biasing the robot toward less crowded regions, maintaining bounded suboptimality relative to the global plan, ensuring adherence to the robot’s minimum turning radius, and preserving kinematic feasibility. Additionally, the pseudo-force cost function is minimized using sequential least squares programming (SLSQP), enabling efficient real-time replanning. In simulation and on a real mobile robot, our method achieves more than 97.9% success rates at low computational cost, with average replanning times ranging between 0.06 and 0.14 s. We benchmarked our results with other prominent sampling-based approaches documented in the contemporary literature. We envisage that the developed path replanning technique holds promising prospects in applications involving manufacturing, healthcare, maritime, and defense, wherein human–robot interaction is essential.
Embodied intelligence requires an agent to complete autonomous navigation in continuous environments following natural language instructions. Existing methods typically output mid-level natural language actions or low-level control commands. However, there is a lack of effective coordination between high-level reasoning and continuous control, which leaves room for improvement in decision interpretability and control stability. To address these issues, we propose the VLN-COT framework, which integrates vision-language navigation models with explicit chain-of-thought reasoning. This framework introduces a generative reasoning mechanism to enhance decision interpretability through structured reasoning processes. It also adopts a dual-modal action representation that simultaneously predicts mid-level natural language actions and low-level continuous control commands, achieving joint optimization between high-level reasoning and low-level execution. During training, we combine supervised fine-tuning with GRPO-based reinforcement learning and design three complementary reward functions to obtain a stable and interpretable navigation decisions. Experimental results on the VLN-CE benchmark show that VLN-COT achieves significant improvements in key metrics such as success rate and path efficiency, validating the effectiveness of explicit reasoning and dual-modal action representation in continuous vision-language navigation.
A tendon-driven hybrid rigid-flexible manipulator with orthogonal hinges is presented for precision spraying applications in unstructured environments. The proposed architecture integrates waterproof electronics and simplified cable routing to enhance environmental adaptability while reducing system inertia. The primary contribution of this work is an integrated mechanical-algorithmic framework. Within this framework, we adapt a chaos-enhanced Particle Swarm Optimization (CE-PSO) algorithm, utilizing established chaotic initialization and adaptive parameter adjustment techniques, to resolve redundant inverse kinematics with 1.28% mean positioning error. This solver is constrained by a practical hierarchical classification strategy that mitigates joint coupling effects. Standard fuzzy PD control and quintic trajectory planning were employed solely to facilitate the validation of this mechanical-kinematic framework, enabling ±30° bending with under 2.39° joint coordination error. Experimental validation confirms a computation time of 0.85s for the inverse kinematics solution, which is suitable for quasi-static or slow-motion tasks such as precision spraying, but may be limiting for highly dynamic operations requiring faster update rates. Notably, dynamic operations exhibit ≤4.65mm end-effector oscillations due to structural compliance. Collectively, the integrated mechanical-computational framework provides a hardware and algorithmic foundation for precision tasks in less structured settings. While kinematic accuracy and static stability were validated in lab tests, comprehensive task-level evaluations and formal safety characterizations are required to fully realize its potential for safe human-robot interaction. However, further real-world validation is necessary to fully assess task-level performance such as spray coverage uniformity and human-robot interaction safety.
