Reliable water surface object detection is a fundamental prerequisite for the practical deployment of autonomous marine systems. However, developing robust perception models conventionally relies on prohibitively expensive dense annotations. While sparse supervision offers a highly cost-effective alternative, existing models suffer catastrophic performance degradation in harsh marine environments. Intense water reflections and dynamic waves exacerbate the scarcity of supervisory signals, leading to severe geometric manifold collapse and the bootstrapping dilemma during pseudo-label mining. To overcome these critical engineering bottlenecks, AquaMine, a sparsely-supervised contrastive instance mining framework, is proposed. The framework integrates a Multi-Class Contrastive Learning module to reconstruct the noise-degraded feature manifold through multi-class parallel interactions, effectively decoupling maritime targets from complex background interference. Building upon this robust representation, a Feature-based Instance Mining module spontaneously extracts latent feature-level pseudo-labels via underlying geometric similarity, successfully bypassing the unreliable predictions of initial detectors. Subsequently, an Intra-Class Feature Calibration module rigorously calibrates these mined signals through strict intra-class alignment against verified ground-truth anchors. Extensive evaluations on the large-scale WaterScenes dataset demonstrate that utilizing merely a 10\% annotation rate, AquaMine achieves a perception efficacy equivalent to 91.7\% of the fully-supervised baseline.

Humanoid robots are increasingly deployed in diverse industrial sectors, including inspection and disaster rescue. However, dynamic maritime environments characterized by wave excitation present significant challenges, as complex six-degree-of-freedom non-inertial motions invalidate the fundamental stationary ground assumption inherent in traditional control methods. This research introduces a composite control framework that integrates an explicit feedforward compensation mechanism with an augmented observation space to neutralize inertial impacts. The proposed methodology employs second-order Taylor expansion to predict platform acceleration trends for preemptive torque generation. Furthermore, physical priors involving relative center-of-mass positions and binary foot contact states are incorporated to eliminate perception ambiguity. Experimental results indicate that embedding analytical mechanical models into the reinforcement learning loop substantially enhances robustness against intense perturbations. The proposed framework achieves a survival capacity three times higher than fully implicit models, offering a stable technical solution for humanoid balance control within non-inertial frames.

Performance degradation caused by temporal-difference signal fluctuations and the inflexibility of fixed proximal constraints remains a key challenge in policy optimization algorithms. To address these issues, this article develops a Coherence-Adaptive Proximal Optimization (CAPO) method. We first derive a temporal-coherence-aware advantage estimation mechanism by measuring the directional consistency of temporal-difference residuals within a local time window. Based on this mechanism, short-horizon and long-horizon advantage estimates are adaptively integrated to provide a more stable and temporally reliable policy improvement signal. CAPO is then instantiated within the PPO framework by incorporating the temporal coherence mechanism into both advantage construction and proximal policy updates. The proposed method preserves the stable optimization structure of PPO while adaptively adjusting the clipping range according to sample-wise temporal coherence. In this way, CAPO allows more sufficient policy improvement when the learning signal is temporally consistent and imposes more conservative updates when the signal is unstable. Experiments on six representative OpenAI Gym control tasks show that CAPO achieves better or comparable performance than standard PPO in most benchmark environments, with statistically reliable improvements in three tasks and comparable final performance in two tasks. Meanwhile, the results also show that CAPO does not uniformly outperform PPO in all environments, indicating that its advantage is task-dependent. Overall, CAPO improves training stability and convergence behavior in several representative tasks, especially when local temporal feedback provides reliable policy-improvement information.

Performance degradation caused by temporal-difference signal fluctuations and the inflexibility of fixed proximal constraints remains a key challenge in policy optimization algorithms. To address these issues, this article develops a Coherence-Adaptive Proximal Optimization (CAPO) method. We first derive a temporal-coherence-aware advantage estimation mechanism by measuring the directional consistency of temporal-difference residuals within a local time window. Based on this mechanism, short-horizon and long-horizon advantage estimates are adaptively integrated to provide a more stable and temporally reliable policy improvement signal. CAPO is then instantiated within the PPO framework by incorporating the temporal coherence mechanism into both advantage construction and proximal policy updates. The proposed method preserves the stable optimization structure of PPO while adaptively adjusting the clipping range according to sample-wise temporal coherence. In this way, CAPO allows more sufficient policy improvement when the learning signal is temporally consistent and imposes more conservative updates when the signal is unstable. Experiments on several representative OpenAI Gym control tasks show that CAPOachieves better or comparable performance than standard PPO in most benchmark environments, with improved training stability and convergence behavior, especially in tasks where local temporal feedback provides reliable policy improvement information.

This paper addresses a method to obtain an accurate 3D human body model and a photorealistic free-view image of an arbitrary person from a monocular camera video. Recent works has shown that it is possible to reconstruct a human model at a level of detail from a single image. However, inferring a complete 3D human model from a network model will be ill-posed if rely on a single photograph of a person. In order to reasonably infer the 3D human model, we propose method based on implicit field representation to integrate the information of video frames by a set of structured latent code. The core of our method is to construct the implicit field by relatively sparse structured latent code. Meanwhile, align the vertices of the parametric human model and structured latent code to the same coordinate system. Extensive experimental results on monocular datasets demonstrate the effectiveness of our approach in generating accurate 3D human models. Our method utilizes a monocular camera to obtain a 3D model which enables consumers create their personality digital model.