To address the limited adaptability of conventional microrobot navigation methods in dynamic and cluttered environments, we propose a learning-based framework for robust real-world microrobot navigation and dynamic target tracking. Our approach utilizes a spatial-temporal transformer reinforcement learning (STTRL) model combined with a deterministic velocity obstacle (DVO) mechanism, which effectively processes observation histories and virtual LiDAR scans to predict optimal control actions. By training the model through large-scale reinforcement learning in randomized simulation environments, we enable the system to extract and utilize contextual information, facilitating adaptive behavior in previously unseen scenarios. Experimental results demonstrate that our framework achieves superior navigation agility with zero-shot transfer capability, highlighting its potential for critical autonomous cell and small creature manipulation tasks.

This paper presents a novel magnetic soft microrobot, controlled by a simplified three-coil electromagnetic platform, designed to simultaneously achieve dexterous cell grasping and transportation. To ensure high-precision manipulation, we established a quantitative bending deformation model for the soft microgripper and developed a robust trajectory tracking strategy combining an extended Kalman filter (EKF) with model predictive control (MPC). Experimental results demonstrate that the proposed EKF-MPC algorithm achieves highly accurate path tracking with a mean absolute error of less than 0.155 mm. Ultimately, the microrobot successfully performed automated pick-and-place operations on zebrafish embryonic cells, achieving a minimal release error of 0.067 mm and highlighting its significant potential for advanced in vitro and in vivo cell manipulation.

This paper presents a novel magnetically actuated soft microjoint, fabricated from programmable magnetic materials, to significantly enhance the dexterity of robotic cell micromanipulation. To address the nonlinear deformation and time-varying disturbances during physical interaction, we established a quantitative deformation model and developed a sliding mode controller integrated with a radial basis function neural network (DMRSMC). Experimental evaluations demonstrated that the microjoint achieves a rapid response time of 16 ms and a maximum deflection angle of 22.16°, with the DMRSMC ensuring highly precise and robust angular tracking compared to conventional PID and SMC controllers. Finally, by integrating the microjoint with a soft micropipette, the system successfully performed dexterous grasping and posture adjustment of zebrafish embryos, achieving a 97% success rate and a minimal orientation error of 1.36°.

To address the compounding errors prevalent in long-horizon dexterous robotic cell micromanipulation tasks, we propose a novel skill information representation imitation learning (SIRIL) framework. Our approach utilizes a VQ-GAN encoder to extract discrete latent codes from expert demonstration videos, while an autoregressive transformer models their temporal distribution to quantify skill information through log-likelihood estimation. By leveraging this log-likelihood as a dynamic safety constraint during inference, the algorithm effectively filters out unsafe actions and mitigates compounding errors over extended execution sequences. Real-world physical experiments demonstrating the complex stripping of zebrafish embryonic cell membranes reveal that our SIRIL method achieves an average subtask accuracy of 86.7% and a final success rate of 64.7%, significantly outperforming existing baseline algorithms.

This work presents a novel imitation-enhanced reinforcement learning framework designed to enable robots to learn dexterous, multi-end-effector cell micromanipulation skills directly from expert demonstration videos. The system utilizes a multi-task observation (MTO) network to extract spatiotemporal trajectories of the end-effectors and deformable cells, alongside a task-parameterised hidden Markov model (THMM) to capture and enforce the implicit constraints of expert actions. To balance manipulation safety and dexterity, we introduce an imitation learning optimisation-based soft actor-critic (ILOSAC) algorithm that facilitates robust skill acquisition through a combination of guided demonstration and autonomous exploration. Real-world physical experiments demonstrate that this automated approach significantly reduces operation time and minimizes cellular deformation compared to current algorithms and manual teleoperation.

To advance automated cell micromanipulation and microassembly, we propose a novel two-finger robotic microgripper fabricated from magnetic programmable soft materials. By establishing an analytical model based on beam deformation theory and magnetic field modeling, we introduce a feedback adaptive grasping strategy (FAGS) that utilizes visual feedback to dynamically regulate the gripping process. Driven by a remote magnetic field, this soft microgripper demonstrates robust, shape- and medium-independent grasping capabilities across various complex microscopic environments. Furthermore, by integrating the microgripper into a multi-degree-of-freedom micromanipulation system, we successfully performed precise, low-damage automated tasks including the grasping, transportation, and posture control of zebrafish embryonic cells.

This paper introduces a Scheduling Adaptive Imitation Learning (SAIL) algorithm to overcome the compounding errors and coordination challenges inherent in long-horizon dexterous robotic micromanipulation tasks, such as cell membrane stripping. The proposed framework utilizes an adaptive scheduler to dynamically switch between a global planning network for generating sparse action sequences during simple tasks, and an interactive planning network for dense, single-step cooperative maneuvers during complex interactions. To facilitate robust learning, we constructed a dual-arm micromanipulation system integrating visual and force feedback, while augmenting expert demonstration data through the SOFA physics simulation platform. Real-world physical experiments on zebrafish embryonic cells demonstrate that SAIL successfully executes delicate membrane stripping surgery with an average accuracy of 84.4%, significantly outperforming existing state-of-the-art imitation learning methods.

This project introduces an automated robotic micromanipulation framework designed to intelligently and precisely dissect intact single cells directly from complex tissue sections. To ensure accurate visual feedback under microscopy, we propose an attention mechanism improved (AMI) neural network that robustly detects and tracks the micro-scale needle tip. The system further integrates marker-free hand-eye calibration, trajectory optimization, and model predictive control (MPC) to seamlessly guide the end-effector along the designated cutting path. Experimental validation on paraffin tissue sections confirms the framework's reliability, demonstrating successful autonomous single-cell extraction with an exceptionally low error margin of under 0.61 μm.

We introduce Dexterous Wristed Robotic Kit(DWRK), designed for autonomous robot-assisted minimally invasive surgery. DWRK provides a reliable modular platform with high dexterity for performing surgical manipulation tasks, enabling data-driven learning approaches toward full surgical autonomy. We implement several benchmark imitation learning algorithms based on this platform.

We designed an autonomous robot for the 2021 ASABE Student Robotics Challenge (Advanced Division) to perform precise agricultural tasks, including the identification and mapping of strawberry plant health statuses. Our system integrated autonomous navigation and advanced computer vision to ensure accurate detection of plant conditions and reliable mapping within complex field environments.
As the team captain, I led the project and was responsible for developing the image recognition modules to identify plant health indicators effectively. Furthermore, I designed and implemented the LiDAR-based navigation and motion planning control algorithms to ensure the robot's precise and stable autonomous operation.

We developed an autonomous robot for the 2021 ASABE Student Robotics Challenge, aiming to achieve accurate identification and mapping of plant health statuses within complex agricultural environments. Our system integrates advanced vision sensing and autonomous control to perform high-precision patrol tasks, successfully securing a perfect score in the competition.
In this project, I was primarily responsible for developing the image recognition modules, which enabled the robust detection of plant components such as leaves and flowers. Additionally, I designed and implemented the navigation control algorithms, ensuring the robot's stable and efficient motion during autonomous field operations.