Zongcai Tan, Dandan Zhang
Optical tweezers (OT) offer exceptional precision for micromanipulation in biomedical applications, achieving submicron accuracy. However, controlling conventional multi-trap OT to perform cooperative manipulation of multiple complex-shaped microrobots in dynamic environments is a significant challenge. To address this, we introduce Interactive OT Gym, a reinforcement learning (RL)-based simulation platform designed for OT-driven microrobotics. Our platform supports complex physical field simulations and integrates haptic feedback interfaces, RL modules, and context-aware shared control strategies specifically tailored for OT-driven microrobots in cooperative biological object manipulation tasks. This integration allows for an adaptive mix of manual and autonomous control, enabling seamless transitions between human input and autonomous operation. We evaluated the platform's effectiveness using a cell manipulation task. The experimental results show that our shared control system significantly enhances micromanipulation performance, reducing task completion time by approximately 67% compared to using pure human or RL control alone and achieving a 100% success rate. With its high fidelity, interactivity, low cost, and rapid simulation capabilities, Interactive OT Gym serves as a user-friendly environment for training and testing advanced interactive OT-driven micromanipulation systems and control algorithms.
Code is available here:
https://github.com/Zongcai23/ICRA2025-OT-Gym.git
The primary challenge in optical tweezers (OT)-driven microrobotics lies in controlling complex-shaped microrobots for cooperative manipulation in dynamic, biological environments. Existing platforms either lack the adaptability required for OT environments or do not integrate human input effectively with autonomous control. Moreover, the lack of simulation platforms and effective shared control strategies hinders progress in optimizing micromanipulation tasks, especially in delicate biological settings.
The Interactive OT Gym is a distributed simulation platform for OT-driven microrobot manipulation, consisting of three modules: a high-fidelity simulator, an RL-based autonomous control system, and a manual control interface with haptic feedback. These modules operate on separate devices, coordinated by ROS, to ensure efficient parallel execution.
We propose a progressive training strategy combining A* path planning with RL-driven speed control for autonomous microrobot navigation. A* handles long-range path planning efficiently, while RL adapts the robot's speed in real-time, addressing challenges like Brownian motion and limited optical trap range. This hybrid approach reduces computational costs and improves training efficiency. The RL agent uses a reward function balancing contact force, collision avoidance, and speed control, ensuring safe and efficient navigation. Training is conducted with an epsilon-greedy strategy to balance exploration and exploitation.
A shared control strategy adjusts the balance between human input and automated navigation based on real-time distance to dynamic obstacles. The system offers three modes: fine-tuning, balanced, and autonomous. The control formula dynamically adjusts the weight parameter to ensure precise manipulation while improving automation efficiency, providing a safe solution for OT-driven microrobot tasks in complex environments.
The B-spline smoothed path outperformed manual and A* paths in path length, curvature, and smoothness, with lower noise and superior precision, making it ideal for high-precision path planning.
The RL training effectively optimized microrobot speed control, balancing navigation efficiency and safety. After around 700 episodes, the model reduced contact forces and kept the robot within 0.06 μm of the trap center, improving stability and safety. This demonstrates the model’s potential for stable, efficient control in complex micromanipulation tasks.
Experiments validated the performance of the OT shared micromanipulation system in task efficiency and user experience, comparing manual, autonomous, and shared control. The shared control system, combining RL-based automation with human input, significantly reduced task completion time, improved success rates, and lowered user workload. It outperformed both manual and autonomous modes, with users reporting higher satisfaction and reduced mental and physical demands.
send an email to jt2923@ic.ac.uk