This paper proposes a real-time trajectory planning method using nonlinear optimization and whole-body control for cooperative transport with heterogeneous mobile manipulators. Unlike conventional approaches that rely on passive path following, the proposed method enables the follower robot to actively generate its own trajectory by laterally offsetting the leader’s path and refining it through nonlinear optimization. The optimization process simultaneously considers inter-robot distance maintenance, geometric consistency, and obstacle avoidance in cluttered environments. The optimized trajectory is executed through a generalized hierarchical control framework that integrates impedance-based coordination between the mobile base and manipulator. Experimental validation with two mobile manipulators—, featuring different locomotion and control characteristics—, demonstrates that the system maintains a consistent end-effector distance while adapting to environmental changes. The proposed method achieves high-fidelity trajectory tracking (average error: 0.037 m, standard deviation: 0.039 m) and maintains inter-robot spacing within a 1.05–1.65 m range for 96.09% of the task duration. The results demonstrate the framework’s relevance to real-world cooperative transport situations, emphasizing the need for adaptability, precision, and inter-robot coordination.

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.