Tiefeng Li

Task decomposition is crucial for enabling robots to master long-horizon compositional environments. Towards this goal, skill primitive learning offers a promising approach. However, solely relying on a single inferred skill abstraction creates an expression bottleneck, rendering it insufficient to capture the complex, heterogeneous kinematic behavior combinations required by tasks even with simple semantic descriptions. To address this, we propose SkillAttention: a framework that dynamically synthesizes context-aware latent skills to ensure precise alignment with task-specific kinematic requirements. These synthesized skills, jointly conditioned with the robot's dynamic state, serve as guidance for a downstream diffusion module to generate coherent motion trajectories tailored to the current task execution status. Additionally, we introduce a language-primed staged fusion module that reinforces the guiding role of textual instructions, effectively mitigating instruction-following failures caused by modality imbalance. Extensive experiments on compositional long-horizon benchmark LOReL demonstrate that SkillAttention achieves significant improvements in both success rates and skill learning efficiency. Transcending simulation-centric state-of-the-art models, our framework successfully bridges the sim-to-real gap, maintaining consistent performance in real-world precise manipulation tasks that are significantly more challenging than their simulated counterparts. Code is available at https://github.com/is-aHuan/SkillAttention.

High hydrostatic pressure alters mechanical behaviors of elastomers by changing molecular mobility and chain confinement. Previous experimental studies have characterized this phenomenon using pressure-dependent relaxation modulus measured under fixed deformation. However, these phenomenological descriptions lack a chain-level physical interpretation and are difficult to generalize to finite deformations and complex loading scenarios. In this work, we extend a previously developed pressure-dependent hyperelastic model by incorporating a free-volume-motivated viscoelastic formulation. The viscoelastic stress is further decomposed into contributions from contour-length relaxation and disentanglement. The pressure dependence of relaxation times is formulated by accounting for the effects of free volume on segmental friction and chain confinement. In situ volumetric compression, rate-dependent uniaxial tension, relaxation, and multistep loading-relaxation tests were conducted on VHB 4910 at hydrostatic pressures ranging from 0 to 90 MPa. The experiments show that hydrostatic pressure prolongs stress relaxation and elevates the stress response. The model reproduces these trends and predicts the response under different loading conditions. The predicted pressure dependence of relaxation kinetics is also consistent with trends reported for other polymeric materials. This framework can be applied for the design of elastomer-based underwater acoustic metamaterials and soft robots operating under extreme hydrostatic pressure.

Soft robots have been developed for underwater tasks owing to their compliant, damage-tolerant bodies, which foster an expectation that soft bodies are inherently robust in aquatic environments. However, most reported systems operate under moderate or quasi-static conditions, and overlook a critical vulnerability—when cavity-rich soft structures are subjected to underwater shocks, the embedded cavities collapse violently, triggering a phenomenon we term ‘soft-matter implosion’. This process generates internal pressures several times higher than the incident pressure and can severely damage onboard electronics. Using a simplified cavity-containing structure, we combine underwater impact experiments and finite element simulations to reveal the mechanism of soft-matter implosion and quantify the peak pressure under different parameters. These results inform a function-structure co-design framework that suppresses implosion while preserving overall flexibility. Guided by this framework, we built a rigid-casing-free, fully integrated manta‑ray‑inspired soft robot that survives single impacts up to 49.8 MPa and sequential impacts at 31.5 MPa, and resumes flapping locomotion after impact. This work establishes an impact‑resistant soft robot platform and provides a generalizable design strategy for durable, flexible machines operating in challenging underwater environments.

Rod-driven parallel continuum robots couple actuation, deformation, and external loading, making motion prediction and interaction sensing interdependent. This paper presents a unified framework for kinetostatic modeling and tip-load sensing in a three-rod parallel continuum robot. A constrained energy formulation relates rod actuation lengths, rod strains, platform pose, and external wrench; task-specific solvers use it for forward deformation, inverse path computation, and load estimation. The finite-element comparison yields a mean tip-position difference of 1.1 mm. Physical experiments give mean relative tip-position errors of 1.16\% without load and 1.69\% under a 500 g load. Planar force-magnitude errors are 4.61\% and 6.23\% for the two tested loading directions. The forward solver averages approximately 30 ms per case, and the load-estimation solver averages approximately 36 ms per solve in the existing computational benchmarks. Estimated loads support recovery of line and circular trajectories after load changes. These solver times support real-time load sensing in the demonstrated disturbance-response workflow, without a distal force/torque sensor.

Untethered soft robots are governed by habitat‐specific constraints because operation without physical tethers requires their compliant structures, actuation mechanisms, and energy‐delivery pathways to remain compatible with the environment. Adopting a habitat‐centric framework, this review compares untethered soft robots across terrestrial, aquatic, aerial, and in vivo environments. On terrestrial substrates, design strategies are shaped by locomotion‐specific demands, including contact modulation in crawling, energy storage and release in jumping, switchable adhesion in climbing, and granular fluidization in burrowing. In aquatic habitats, the dominant constraint shifts with depth, from capillary support at the air‐water interface, to hydrodynamic drag in shallow water, and to hydrostatic pressure in the deep sea. In aerial environments, strict mass and power limits favor actively flapping insect‐scale robots, buoyancy‐assisted vehicles, and passively dispersed aeroelastic structures. In vivo, miniaturization and biocompatibility confine the viable design space to magnetic, acoustic, and biohybrid actuation. A multi‐environment comparison shows that no single actuation modality is universally optimal. Each habitat filters actuation strategies by the compatibility between energy transmission and the local medium. Material longevity, onboard energy autonomy, multi‐environment transition, extraterrestrial operation, and perception and feedback control are identified as cross‐cutting challenges whose resolution is required to translate laboratory demonstrations into real‐world deployment.

概 要 柔性电子以柔性材料、 可变形结构和集成化信号处理为基础, 能够在曲面、 运动以及复杂环境中实现感知与信息交互。 近年来, 柔性电子在可穿戴健康监测、 人机交互、 软体机器人和智能制造等领域发展迅速, 并且逐步从单一器件性能优化拓展到材料、 结构、 传感机制、 算法和系统应用的协同设计。 本专辑收集了本领域具有代表性的研究成果, 介绍了柔性水下机器人、 离子电子压力传感器、 可调应变传感器、 柔性磁力传感器、 激光诱导石墨烯脉搏传感阵列以及温湿度解耦传感器等方面的研究进展, 展示了柔性电子在高灵敏感知、 信号解耦、 稳定封装、 智能识别和软体系统控制中的应用潜力。 希望本期内容能够帮助读者了解柔性电子领域的研究趋势, 推动柔性器件与人工智能、 软体机器人及智能健康技术的进一步融合, 为柔性电子系统的可靠化、 集成化和工程化发展提供参考。

This paper presents an integrated framework for the design and control of a morphing wing for fixed-wing Unmanned Aerial Vehicles (UAVs), combining a hybrid actuation system with a deep reinforcement learning (RL) controller. A novel hybrid mechanism using Shape Memory Alloy (SMA) wires for continuous trailing edge camber control and a servo motor for leading edge droop is proposed. A high-fidelity coupled Computational Fluid Dynamics (CFD) and Computational Structural Dynamics (CSD) model is developed to simulate aeroelastic responses, from which a Reduced-Order Model (ROM) is derived for efficient RL training. The control problem is formulated as a Markov Decision Process (MDP), and a Deep Deterministic Policy Gradient (DDPG) agent is trained to maximize the lift-to-drag ratio (L/D) while minimizing control energy. Simulation results demonstrate that the RL controller achieves an average L/D improvement of over 19% across various flight conditions, effectively delays stall, maintains structural integrity with skin strains below 0.4%, and reduces energy consumption by 8–16% compared to a rule-based controller. The study validates the effectiveness of RL in real-time multidisciplinary optimization for morphing wing systems.

Cavitation is an undesirable phenomenon in hydrodynamic systems, leading to surface erosion, efficiency degradation, and intense acoustic emissions. Existing research efforts have primarily focused on mitigating the impact of cavitation, yet none have addressed the suppression of its initiation. Here, we propose a hydrogel coating strategy that fundamentally raises the energy barrier for heterogeneous nucleation, thereby suppressing the onset of cavitation. The coating simultaneously retains the capability to dissipate impact energy upon bubble collapse. When applied uniformly to hydrodynamic surfaces, it exhibits robust adhesion, durability, and anti-swelling performance. Under representative conditions, the hydrogel coating maintains a thrust enhancement of approximately 3.4% under cavitating flow and elevates the onset speeds for both incipient and developed cavitation, resulting in a 31.3% increase in maximum thrust under cavitation-free operation and a 27.8% increase in maximum thrust before the onset of developed cavitation. Additionally, the coating substantially attenuates broadband noise, achieving a maximum reduction of 12.8 dB during developed cavitation. This hydrogel-based approach offers a new direction for cavitation suppression, enabling enhanced efficiency, stability, and noise reduction across diverse hydrodynamic systems.

Soft fluid actuators (SFAs) are crucial in the design of underwater systems, providing essential deformation capabilities for various applications. However, their power sources are typically rigid and bulky mechanical pumps, which significantly restrict their flexibility and environmental adaptability. To address this issue, we propose a magnetically enhanced, small-scale, lightweight dielectric elastomer (DE) pump for underwater applications. This novel pump achieves a maximum underwater driving pressure of 2.18 kPa and a maximum average flow rate of 81 ml/min. A dynamic model, based on the uniform deformation of the DE film, was developed to describe the driving behavior of the magnet-DE actuator. Utilizing the proposed DE pump, we designed balloon-type and gripper-type flexible actuators, capable of underwater grasping complex-shaped objects, thereby verifying the pump’s efficacy in driving SFAs underwater.