In physical human-robot interaction (pHRI), ensuring human safety in all tasks conducted by the robot is crucial. Traditional compliance control strategies, such as admittance and impedance control, often lead to unpredictable robot behavior due to incidents like contact loss or unexpected external forces, which can cause significant harm to humans. To overcome these limitations, this study introduces a multi-task energy-aware impedance controller for kinematically redundant robots. This controller extends the energy-aware impedance control strategy, which ensures the passivity and safety of a single task using a virtual global energy tank, to kinematically redundant robots performing multiple tasks. The proposed controller effectively regulates the power flow of all tasks performed by the robot through a single global energy tank, ensuring the safety and passivity of the tasks. Experimental results in a shared environment, where external forces are simultaneously applied to the end-effector and the third joint of the Franka Emika Panda, showed that the robot's energy and power, as well as the power of all tasks, consistently remained within predefined thresholds. Additionally, when comparing the proposed controllers with controller that do not consider null space projection in the power regulation stage and controller that do not regulate the robot's power, our approach effectively managed the robot's energy and power and the power of all tasks, ensuring passivity and enhanced safety.

Gait is intricately linked to human health, with key gait indexes showing strong correlations. To overcome limitations of flat-surface-focused evaluations, we propose a Modified Gait Index (MGI) that adapts to diverse terrains such as ramps and stairs. Gait data were collected from 15 healthy adults using motion capture, IMU (Inertial Measurement Unit), EMG (Electromyography), and force plates. For each terrain, principal component analysis (PCA) identified key contributing features based on loading vectors and contribution scores. MGI formulas were derived by combining these features with clinically relevant variables, including stance/swing ratio and joint flexion angles. Participants were divided into training and test groups based on PCA similarity, and generalization was evaluated using interquartile range-based normal thresholds. Across all terrains, the test group’s average GI values remained within the training-defined normal range: 1.05 - 2.82 for ramp ascent, 0.71 - 4.02 for ramp descent, 1.68 - 2.86 for stair ascent, 1.15 - 3.99 for stair descent. Although some test values approached quartile boundaries, no outliers were observed. These results demonstrate the MGI’s terrain-adaptive consistency and generalizability, supporting its application in exoskeleton control strategies for users with subclinical gait deviations or muscular overload.

On-device Human Activity Recognition (HAR) requires balancing accuracy and deployment efficiency on constrained hardware. We present Lightweight Human Activity Recognition (L-HAR), a controlled comparison of Baseline and Lightweight Temporal Convolutional Network (TCN), Transformer, and Long Short-Term Memory (LSTM) models using data sampled at $100\mathrm{Hz}$ from three inertial measurement units (IMUs) worn by one participant. Evaluations covered classification, model footprint, multiply–accumulate operations (MACs), offline CPU latency, Desktop/Raspberry Pi streaming, 8-bit integer (INT8) quantization, and software-estimated Raspberry Pi energy efficiency. Lightweighting reduced parameters by up to $95.5\%$, model size by $94.0\%$, and MACs by $87.6$–$95.5\%$, with accuracy reductions of $0.8$–$2.2$ percentage points (pp) and sub-millisecond offline latency for all Lightweight models. Baseline and Lightweight models sustained approximately $100\mathrm{Hz}$ message/inference rates on Desktop, whereas no Raspberry Pi configuration reached $100\mathrm{Hz}$ inference. Among Lightweight models, Transformer achieved $9.381\mathrm{ms}$ End-to-End (E2E) latency and $66.925\mathrm{Hz}$ inference. INT8 quantization changed accuracy and F1-score by less than $0.13$ pp; Quantized TCN achieved the best Raspberry Pi streaming result ($9.066\mathrm{ms}$ E2E; $69.928\mathrm{Hz}$) with an estimated $49.933\mathrm{mJ}$ per inference. These results demonstrate architecture- and backend-dependent deployment behavior and the need for direct target-platform evaluation.