Compact V-shaped in-pipe robots are attractive for small-diameter inspection because they combine wheeled efficiency with multi-point wall support. However, three-dimensional fittings such as an out-of-plane double elbow (OPDE) expose a fundamental locomotion problem: axial propulsion and body rolling must be available at the same time, while their wheel-wall contacts can transmit unintended motion between the two functions. This paper presents SPIKR-1.0, a two-motor V-shaped in-pipe robot designed to suppress this propulsion-rolling crosstalk at the structural level. The robot separates the V-arm joint axis from the spherical-wheel axis, assigning all-wheel-drive omni wheels to axial propulsion and a centrally actuated spherical wheel to rolling reorientation. To make this design principle analyzable, a geometric reduction relates the spherical-wheel mounting layout to the contact angle, a passive-preload model predicts the contact normal loads and friction capacity, and a contact-limited axial-roll transmission model explains how rolling traction is divided into useful circumferential motion and undesired axial leakage. The framework is evaluated through effective friction measurements, numerical contact-capacity studies, vertical-pipe spherical-wheel-radius ablation experiments, and repeated teleoperated traversal tests on a multi-material pipe testbed containing an OPDE. The results show that contact-angle-guided structural design can reduce rolling-induced propulsion disturbance while preserving posture-adaptive passability in compact in-pipe robots.
Zero-Shot Anomaly Segmentation (ZSAS) detects anomalous regions in unseen classes without training samples, and is valuable in data-restricted industrial and medical domains. However, existing methods based on Contrastive Language-Image Pretraining (CLIP) rely on class-agnostic text prompts that fail to capture class-specific anomaly characteristics, resulting in semantic ambiguity and misalignment with patch-level visual features. To overcome these limitations, this study proposes a novel ZSAS framework that leverages a Multimodal Large Language Model (MLLM) to automatically generate Class-Specific Text Prompts (CSTPs), which provide detailed prompts describing each class’s physical structure, normal patterns, and representative anomaly characteristics. The proposed method strengthens the alignment between CLIP’s visual and textual features, enabling more precise anomaly localization. Moreover, an adapter-based residual connection structure and an alternating learning strategy enable stable training while preserving CLIP’s pretrained knowledge. Experiments on diverse industrial and medical datasets show that the proposed approach attains the highest mean Anomaly Detection (mAD) over the industrial and medical domains and the strongest medical-domain performance, while offering a substantially better accuracy–efficiency trade-off in inference speed and memory. These results confirm that MLLM-generated class-specific prompts effectively enhance generalization for zero-shot anomaly segmentation.
This paper investigates robust model predictive control (RMPC) for discrete-time Markov jump linear systems whose sensor-side state estimates are transmitted through noisy digital channels. A dynamic binary encoder with fractional bit depth is introduced first, where the resolution factor enables sub-integer representation of the scaled residual and reduces the quantization floor of integer-only encoding. A fractional-bit-depth encoding with selective bit protection (FBDE-SP) mechanism is then developed. Its adaptive residual-budget-guided selection rule determines, for each encoded component and transmission instant, the largest admissible contiguous set of low-sensitivity magnitude bits that can remain unprotected, while the sign bit and the remaining protection-critical bits are protected. The resulting quantization and bit-flip effects are converted into a certified componentwise communication-residual set whose budgets are linked explicitly to the robust input margins of the controller. This residual is embedded as a bounded input into an augmented Markov jump model of the plant state and observer error. For an explicitly stated mode-history-dependent multistage policy class that is closed under state-dependent terminal recourse, LMI-based terminal conditions establish recursive feasibility, robust constraint satisfaction, and uniform mean-square boundedness. A tractable tightened-RMPC implementation is used for numerical evaluation. Numerical results on an aircraft-engine benchmark show that the proposed strategy approaches the closed-loop performance of full bit protection in the reported setting while using fewer protected bits.
Spatio-temporal corridors are commonly used for unmanned ground vehicle trajectory planning, but their pre-defined geometric constraints may overly limit drivable space in cluttered terrains, yielding infeasible or high-cost solutions. Inspired by how human drivers selectively use visual cues for decision-making, we propose an attention-enhanced trajectory planning framework to guide where corridors should be contracted or extended. We first design a lightweight recurrent NCP module (with only 22.3k parameters) to identify visual salient features to vehicle control actions. These features are then projected into vehicle coordinates as constraints to corridor optimization. Experimental results show lower traversal costs and faster visual inference than semantic-based methods, while generating corridors in real-time. Our framework’s effectiveness is validated in unstructured terrains with random obstacles.
Task scheduling for lunar communication-navigation integrated constellations involves four coupled decision dimensions, i.e. task selection, task sequencing, coordinated service assets determination, and satellite-to-task assignment, which are tightly coupled due to task priority, limited load for dual function, and inter-satellite coordination. Most existing methods address this NP-hard scheduling problem through rule-based heuristics or reinforcement learning with fixed resource allocation schemes, both of which lead to significant limitations in optimality. In this paper, we develop an end-to-end deep reinforcement learning (DRL) framework for joint task scheduling, aiming to learn an effective joint scheduling policy for the coupled optimization problem. To better match the problem structure, we adopt a Dual-Attention Proximal Policy Optimization framework, in which self-attention and cross-attention are used to encode inter-task resource competition and task-satellite compatibility, respectively. More importantly, the proposed multiple decision heads explicitly incorporate newly designed deferral and dynamic K-selection, enabling proactive congestion control and adaptive quality-resource trade-offs, which together yield improved overall scheduling performance. The proposed scheduling framework generalizes well to constellations with the same resource-sharing structure, including those in which shared payloads concurrently support multiple service types. Numerical simulation for Elliptical Lunar Frozen Orbit satellite constellation shows that the proposed learning algorithm achieves 14.7% higher revenue than the best heuristic baseline, and 21.4% higher than a standard DRL baseline, with the gain confirmed on 20 held-out instances under a common evaluation protocol.
