Most existing research in human pose estimation focuses on predicting joint positions, paying limited attention to recovering the full 6D human pose, which comprises both 3D joint positions and bone orientations. Position-only methods treat joints as independent points, often resulting in structurally implausible poses and increased sensitivity to depth ambiguities—cases where poses share nearly identical joint positions but differ significantly in limb orientations. Incorporating bone orientation information helps enforce geometric consistency, yielding more anatomically plausible skeletal structures. Additionally, many state-of-the-art methods rely on large, computationally expensive models, which limit their applicability in real-time scenarios, such as human–robot collaboration. In this work, we propose STAG-Net, a novel 2D-to-6D lifting network that integrates Graph Convolutional Networks (GCNs), attention mechanisms, and Temporal Convolutional Networks (TCNs). By simultaneously learning joint positions and bone orientations, STAG-Net promotes geometrically consistent skeletal structures while remaining lightweight and computationally efficient. On the Human3.6M benchmark, STAG-Net achieves an MPJPE of 41.8 mm using 243 input frames. In addition, we introduce a lightweight single-frame variant, STG-Net, which achieves 50.8 mm MPJPE while operating in real time at 60 FPS using a single RGB camera. Extensive experiments on multiple large-scale datasets demonstrate the effectiveness and efficiency of the proposed approach.
9 March 2026 in IEEE Robotics and Automation Letters
Driven by significant advancements in structure, sensors, and control algorithms, multi-fingered hand systems have received increasing attention from both academia and industry. While existing research primarily focuses on the interaction static rigid objects manipulation, real-world scenarios often involve dynamic tasks and non-rigid materials. To address these challenges, we developed a multi-fingered manipulation system designed to catch dynamic, non-rigid objects with small-deformation. We decomposed this objective into three components: teleoperation, reinforcement learning-based control, and deformation feedback. This paper presents a 3D vision-based framework that integrates these elements. Experimental results demonstrate that the framework can catch a paper sheet with 5.40 mm deformation and a paper balloon with 12.19 mm deformation. Regarding system latency, the latencies for teleoperation, RL-based control, and deformation feedback are 1.29 ms, 0.65 ms, and 0.47 ms, respectively.
25 March 2026 in IEEE Robotics and Automation Letters
Although many impressive advances have been reported in the table tennis robots field using reinforcement learning method, challenges related to policy complexity and adaptability continue to hinder large-scale deployment and practical applications. In this work, building upon extensive prior studies, we propose several techniques and integrate them into a unified framework that reduces the difficulty of training and deployment while enhancing the human player enjoyment. Specifically, a recursively nested design is introduced in the hierarchical decision-making system, which fully separates high-level decision-making from low-level execution, eliminating the need to train multiple agents at the execution layer. The multi-objective problem is also studied by introducing a tolerance-based weight modulation mechanism, which can balance the landing accuracy with other objectives and adjust the playing strategy to meet diverse goals. Comprehensive experiments demonstrate the effectiveness of the proposed framework and techniques. Notably, the robot achieved up to 18 consecutive rallies with human player, which, to the best of our knowledge, is the longest rally attained with a fixed-base collaborative robot arm. The proposed methods can be readily extended to other racket-sport robots or similar tasks.