Maritime autonomous navigation enables unmanned surface vehicles to perform multiple tasks and can also serve as a driver-assistance system for manned vessels, improving safety and comfort. Compared with on-road scenes, maritime perception faces unique challenges: optical sensors degrade due to reflections and sea fog, while wind and waves cause large platform motions that disrupt perception stability. Therefore, high-quality multi-modal datasets are crucial. Meanwhile, 4D millimeter-wave (mmWave) radar has been successfully applied in on-road scenes due to its all-weather, all-day, and high-resolution sensing capabilities, but its maritime characteristics and waveform design effects remain underexplored. To address these challenges, we present MariSense, a multi-modal dataset for maritime autonomy perception. MariSense is the first to introduce the concept of dynamic waveform 4D mmWave radar into maritime scenes, integrating multi-waveform radars with visible and thermal cameras, LiDAR, marine radar, Automatic Identification System (AIS), and navigation sensors, all with extrinsic calibration. The dataset covers various scenes such as anchorages and fairways under varying weather, time, and sea conditions. We establish benchmarks for object detection, tracking, and odometry, and evaluate baseline methods, highlighting the need for improved maritime perception, multi-modal fusion, and adaptive 4D mmWave radar waveform design. The dataset can be downloaded from https://github.com/ORCAUBOAT/MariSense .
Developing generalist end-to-end autonomous driving systems requires balancing the retention of fundamental driving capabilities with the flexibility to adapt to diverse downstream objectives. Current paradigms struggle to reconcile these requirements. Direct regression methods rely on rigid representations that are prone to catastrophic forgetting during adaptation, while diffusion-based approaches are often limited by entangled policy representations and high computational costs. To bridge this gap, we present Driving On Manifolds (DOM), a novel framework based on the Stochastic Manifold Mapping (SMM) paradigm. DOM mathematically decomposes the driving problem into learning a reversible geometric mapping onto the action manifold and modeling a latent density over policy preferences. This topological separation confers two inherent advantages. First, the learned geometric mapping enables extensive synthetic data generation via latent sampling, allowing DOM to acquire intrinsic safety awareness by internalizing boundary constraints without relying on reactive external simulation. Second, it facilitates capability-preserving adaptation: the system aligns with downstream tasks (e.g., via Reinforcement Learning) by updating only the latent distribution and leveraging the exact likelihood for efficient fine-tuning, while the frozen geometric mapping structurally confines adaptation to the pre-trained manifold, substantially mitigating the degradation of established driving capabilities. Specifically, we instantiate this framework using a Manifold Normalizing Flow, driven by features from ContextFormer, a streamlined encoder designed for efficient environmental feature extraction. Validated on the NAVSIM benchmark, large-scale real-world datasets, and closed-loop real-world experiments, DOM achieves state-of-the-art performance and demonstrates favorable scaling properties, offering a unified and scalable foundation for autonomous driving.
Achieving anthropomorphic agility in robotic systems remains a fundamental challenge, heavily constrained by the mechanical limitations of conventional serial-link actuators. Current joint architectures suffer from an inherent trade-off between torque density, structural compactness, and kinematic singularity, failing to replicate the complex and coupled spherical motion of biological joints. To bridge this gap, this paper proposes a high-dexterity 3-DoF dexterous joint actuator (JAMS) by in-depth mimicking the bionic advantages of the human hip. Unlike traditional decoupled designs, JAMS introduces a load-adaptive hierarchical actuation strategy that synthesizes power from three parallel motors via customized differential bevel gear and spherical gear trains. This architecture enables omnidirectional rotation around a spatially fixed center, effectively resolving the kinematic redundancy and structural bulkiness of serial equivalents. A comprehensive kinematic model is derived to map the complex gear coupling, facilitating the development of an effectively singularity-free motion control algorithm implemented on a custom dual-MCU embedded controller board. Experimental results demonstrate that JAMS achieves an extensive operational space of 360 ° × 250 °×360° and a theoretical rated torque of 21.5 N·m, exhibiting high torque density within a compact volume. Trajectory tracking experiments verify the actuator’s capability to track selected human hip reference trajectories under scaled-duration gait-cycle conditions, with a composite kinematic-fidelity score of 95.1%, demonstrating its fundamental kinematic feasibility. By integrating multi-drive hierarchical output and spherical rotation mechanism, this work presents a validated mechatronic solution that overcomes long-standing engineering bottlenecks, offering a modularized architecture for multi-DoF robotic joint design, with potential relevance to humanoid robots, robotic arms, and future wearable mechanisms.
We investigate the topological structure of the optimal actuation landscape for fully-actuated N-rotor aerial vehicles. By formulating the design problem on the 2N-dimensional product manifold of projective lines (RP2)(N)and minimizing a rotation-invariant Log-Volume isotropy metric, we map how optimal rotor orientations evolve across diverse polyhedral chassis. The results establish that global optimality is strictly bounded by geometric symmetry. While irregular chassis yield discrete, isolated optimal configurations, regular geometries induce a structural phase transition: the optimal space initially collapses onto an N-dimensional Tangent Torus, then systematically reduces to continuous configurations governed by affine phase coordination. These collapses define the "N-5 Scaling Law." For N <= 7, the optimal landscape fundamentally forms exactly K = N-5 disconnected 1D closed loops. For N >= 8, these 1D trajectories expand into core backbones embedded within multi-dimensional flat optimal hypersurfaces. Furthermore, for regular planar geometries, we theoretically unify these trajectories by demonstrating a strict geometric isomorphism to star polygons {N/q} (2 < q < N - 2), yielding exact analytical predictions of the N-5-branched optimal manifolds for arbitrary N. By shifting the classical framework from single and static optimal points to continuous optimal manifolds, this work establishes the mathematical foundations for optimality-preserving morphing. The identified topological branches provide continuous null spaces, introducing to the field the concept of design redundancy, and allowing symmetric multirotor systems to dynamically reconfigure their thrust vectors without degrading their perfectly isotropic control authority.
In precision agriculture, an autonomous weed-removal robot (i.e., mobile manipulator) often faces a combination of challenges including large-scale navigation, repetitive coverage, and fine-grained intervention. These challenges give rise to a distinct energy-efficient long-horizon planning problem: coordinating ordered, repetitive, and heterogeneous tasks under domain-specific temporal and geometric constraints. In precision weeding, this problem is further characterized by (i) an observe-first-and-manipulate-later type temporal constraint and (ii) the potential for substantial energy savings through efficient task collocation. We formalize this challenge and propose a new problem: Repetitive and Clustered-Coupled Active Perception and Manipulation (RC-CAPM) problem for mobile manipulators. We model RC-CAPM as a stochastic renewal process and show, via the Renewal Reward Theorem, that the expected energy consumption per task cycle converges to a well-defined long-run average. We introduce a task-space partitioning method that enumerates all feasible task-multiplexing configurations and their associated probabilities under robot reachability constraints. We then derive a two-stage optimization framework and prove that its solution is optimal and equivalent to that of the original RC-CAPM problem. We evaluate the proposed approach in both simulation and on a physical mobile manipulator performing autonomous weeding. Across all evaluations, RC-CAPM consistently outperforms baseline planning strategies, yielding reductions in path length, number of stops, replanning events, and total energy consumption. In particular, the proposed method achieves an average energy reduction of 36.1% relative to the baseline.
