In humanoid robot control, the inverse kinematics solver constitutes a critical real-time component, as any latency may compromise stable operation. This study presents the development and evaluation of two hybrid methods for solving the real-time inverse kinematics of a six-axis anthropomorphic leg lacking a spherical joint. The absence of this joint makes closed-form analytical solutions not possible, necessitating fast and accurate numerical approaches for real-time applications. To accelerate convergence and reduce the overall number of required iterations, we propose two distinct strategies for generating an initial guess for a Damped Least Squares numerical solver. The first is a novel geometry-based method designed for the specific kinematic constraints and reachable workspace of the leg structure. The second employs a custom neural network trained on forward kinematics data. Comprehensive testing was conducted to evaluate convergence behavior based on iteration count. Execution time measurements obtained directly on a microcontroller showed that from the proposed methods, the quickest one takes an average of 3.24 ms of calculation time with a maximum of 3.9 ms. This validates the real-time viability of both proposed approaches, demonstrating a 37
Low-light imaging conditions introduce signal-dependent Poisson noise that degrades structural detail and compromises the reliability of downstream computer vision algorithms. The present work investigates resource-efficient denoising strategies for preserving morphological image features under non-stationary noise regimes. A class of architectures designated Filter-Guided Hybrid Cascades (FGHC) is proposed and evaluated, in which lightweight deterministic filters supply structural priors to compact neural restoration models. The study systematically compares classical filtering approaches, standalone neural baselines, and four hybrid integration strategies, namely Early Fusion, Residual Addition, Spatial Attention, and Spatial Feature Transform, with respect to restoration fidelity and computational overhead. Experiments conducted on a fluorescence microscopy denoising benchmark demonstrate that the strongest hybrid cascade configurations improve restoration quality relative to isolated baselines. The best-performing overall configuration, consisting of an Anisotropic Diffusion Zhang filter coupled with NAFNet under Spatial Feature Transform modulation, achieves a PSNR gain of 5.46 dB and an SSIM of 0.941. For deployment under resource-constrained conditions, the configuration combining a Gradient-Guided Bilateral Filter with Noise2Detail under Spatial Feature Transform modulation attains the most favorable efficiency trade-off, delivering a PSNR gain of 5.32 dB, an SSIM of 0.938, a throughput of 10.6 frames per second, and a VRAM footprint of 79.4 MB. The reported results indicate that the incorporation of deterministic structural priors into lightweight neural denoising pipelines improves restoration performance while preserving the latency and memory characteristics required for edge and robotic perception deployment.
Effective perception in multi-robot systems depends both on the physical placement of sensors and on their active control during operation. However, existing approaches typically address sensor placement and active sensing separately, neglecting the dependency between hardware configuration and sensing strategies. In this work, we propose a two-stage framework for cooperative active sensing in robot teams equipped with multiple steerable sensors. In the first stage, sensor placement and base orientations are optimized using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) multi-objective genetic algorithm, balancing field-of-view overlap and angular sensing density under platform constraints. In the second stage, cooperative sensor control is formulated as a reinforcement learning problem. Centralized policies trained with three algorithms—Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C), and Soft Actor-Critic (SAC)—learn to rotate sensors within bounded ranges around the optimized orientations to maximize environmental coverage while maintaining visual observation of teammates, with SAC achieving the highest reward improvement of up to 11.5
Control of biomimetic underwater vehicles (BUVs) is complicated by the nonlinearity of hydrodynamics and the lack of accurate analytical models, which limits the application of classical controllers. This paper proposes a control system for a robotic fish with a multi-link caudal fin based on deep reinforcement learning without explicit hydrodynamic modeling. The problem is formalized as a Markov decision process and solved using the Soft Actor-Critic (SAC) algorithm with a multi-component reward function in the Stonefish 2 simulator integrated with ROS2. Experimental results demonstrate that the trained policy achieves a steady-state depth error of 0.07 m with a standard deviation of 0.04 m, a settling time of 22 s, and a maximum velocity of 0.15 m/s while maintaining heading. The results confirm the effectiveness of the proposed approach for developing adaptive control systems for biomimetic underwater vehicles.
Energy efficiency is a central objective in the design of reinforcement learning (RL)-based controllers for legged robots. Yet, the influence of individual gait parameters on energy consumption in learned locomotion policies remains underexplored. This paper presents a systematic sensitivity analysis of the cost of transport (CoT) across an 11-dimensional gait-parameter space for a learned quadrupedal locomotion policy deployed on the Unitree AlienGo platform. Gait commands are varied over predefined ranges, and the resulting CoT is evaluated at multiple locomotion velocities. To ensure experimental validity, we assess the policy’s ability to reliably track the commanded gait parameters. Our analysis reveals that only a subset of five parameters exerts a dominant influence on CoT, with the effect strongly modulated by locomotion speed, while the remaining six parameters contribute negligibly within the examined operating envelope. Furthermore, we show that a data-driven surrogate model–specifically, a multilayer perceptron–can accurately predict CoT directly from commanded gait parameters and velocity. These findings offer practical guidance for reducing the dimensionality of energy-aware gait optimization and enable rapid, model-based screening of efficient locomotion strategies in learned quadruped controllers.
