Zhufeng Shao

Robotic LEGO brick assembly represents a challenging case of rigid precision assembly, requiring simultaneous multi-stud alignment under extremely tight tolerances. Unlike conventional peg-in-hole tasks where contact force diminishes as alignment improves, LEGO's interference-fit mechanism requires sustained and increasing force during a snap-engagement phase. This inverted force–outcome relationship often causes standard reactive policies to prematurely reduce force and fail at the final insertion stage. Furthermore, slight misalignment can generate upward reaction forces along the gripper axis, leading to object slip and in-hand pose drift—failure modes that are difficult to detect using instantaneous observations alone. To address these challenges, we propose a future-aware control framework based on Long Short-Term Memory (LSTM) that predicts short-horizon future end-effector trajectories and incorporates this prediction into the reinforcement learning objective. Instead of relying solely on instantaneous state feedback, the policy is guided by predicted future alignment consistency, enabling it to distinguish between transient contact disturbances and true misalignment. A future-drift penalty is introduced to encourage stable convergence toward insertion while maintaining sufficient contact force during the snap phase. Simulation results demonstrate that the proposed framework significantly improves robustness and success rate in LEGO assembly tasks, particularly in the presence of subtle misalignment and complex contact dynamics.

Since the Third Industrial Revolution, robotics and automated systems have become deeply integrated into industrial production and daily life, fundamentally reshaping production efficiency and product quality

Kirigami-inspired geometries offer a lightweight, bioinspired strategy for friction enhancement and anchoring in soft robotics. This study presents a bioinspired kirigami structure designed to enhance the anchoring performance of soft robotic systems through systematic geometric and actuation parameter optimization. Drawing inspiration from the anisotropic friction mechanisms observed in reptilian scales, we integrated linear, triangular, trapezoidal, and hybrid kirigami cuts onto flexible plastic sheets. A compact 12 V linear actuator enabled cyclic actuation via a custom firmware loop, generating controlled buckling and directional friction for effective locomotion. Through experimental trials, we quantified anchoring efficiency using crawling distance and stride metrics across multiple cut densities and actuation conditions. Among the tested configurations, the triangular kirigami with a 4 × 20 unit density on 100 µm PET exhibited the most effective performance, achieving a stride efficiency of approximately 63% and an average crawling speed of ~47 cm/min under optimized autonomous operation. A theoretical framework combining buckling mechanics and directional friction validated the observed trends. This study establishes a compact, tunable anchoring mechanism for soft robotics, offering strong potential for autonomous exploration in constrained environments.

Parallel robots, including cable-driven parallel robots (CDPRs), are widely used due to their high stiffness, precision, and high dynamic performance. However, their multi-chain closed-loop architecture brings nonlinear, multi-degree-of-freedom coupled motion and sensitivity to geometric errors, which result in significant challenges in their modeling, error compensation, and control. The rise in machine learning technology has provided a promising approach to address these issues by learning complex relationships from data, enabling real-time prediction, compensation, and adaptation. This paper reviews the progress of typical applications of machine learning methods in parallel robots, covering four main areas: kinematic modeling, error compensation, trajectory tracking control, as well as other emerging applications such as design synthesis, motion planning, and CDPR fault diagnosis. The key technologies used, their implementation architecture, technical difficulties solved, performance advantages and applicable scope are summarized. Finally, the review outlines current challenges and future directions. It is proposed that hybrid learning physics modeling, transfer learning, lightweight deployment, and interdisciplinary collaboration will be the key directions for advancing the integration of machine learning and parallel robotic systems.