This project was developed as part of my bachelor's graduation thesis in robotics engineering. It features a soft pneumatic gripper system that uses an AI-based controller to adjust its grasp based on object stiffness, sensed through embedded tactile sensors. The system is designed to operate without predefined configurations or manual adjustments, enabling flexible handling of unknown objects.
This section outlines the hardware design and integration of the AI-based soft gripper system, including the experimental setup, actuator design, and electronic control layout.
The experimental platform includes a soft gripper mounted on a linear positioning system with a lead screw mechanism driven by stepper motors. The base and Z-axis stage are reinforced using aluminum extrusions and laser-cut acrylic to ensure mechanical stability and minimize sensor interference during testing. The system designed for repeatable, stable trials during data collection and model validation.
The soft gripper uses DragonSkin 20-based PneuNet bending actuators with embedded bending and force sensors. These actuators are cast using 3D-printed molds with pre-positioned sensors.
The system is powered and controlled via a modular electronics setup, consisting of:
- 12V Air Pump for pneumatic actuation
- L298N Motor Driver for pump control
- TMC2208 Drivers for stepper motor movement
- Voltage & Current Sensors for monitoring actuation load
- Arduino UNO Microcontroller handling PWM generation and sensor data acquisition
A force estimation model was implemented to calculate the applied grasping force based on bending and PWM input. The model was constructed using simulation data from an FEA compression test and validated with a real-world physical compression setup. This reference data was used to map bending behavior and actuation effort to estimated contact force during grasping.
FEA Compression Test
Experimental Compression Test
These demonstrations illustrate key behaviors of the soft gripper during real-time operation, highlighting the system’s ability to control applied force and react to interaction changes.
Grasping.Rubber.Ball.mp4
Grasping a soft rubber ball using real-time force estimation
This test shows the soft gripper grasping a rubber ball, with force feedback used to modulate actuation effort. The system tracks bending and force in real-time, stopping when the AI controller predicts that the desired grasp force has been reached.
Grasping.Sample.mp4
Grasping a TPU sample and responding to object removal mid-grasp
In this demo, the gripper begins to grasp a rigid sample. Midway through, the object is manually removed to simulate unexpected release. The system detects the loss of contact through a sudden drop in estimated force, automatically stops actuation, and enters a safe state as indicated by the red system status.