PROJECT TOPICComputer Vision & Embedded AI
Semantic Segmentation Gripper
A manipulation research concept investigating lightweight semantic segmentation networks running on wrist-mounted cameras to detect irregular target objects and compute optimal grasp angles.
Pixel-Level Perception for Robotic Manipulation
- ✓Overcomes clutter by isolating individual object masks down to pixel boundaries.
- ✓Computes surface normal vectors to position gripper jaws perpendicular to object faces.
- ✓Features eye-in-hand camera mounting to eliminate line-of-sight occlusion.
The Engineering Challenge
Problem / Objective
Standard bounding-box object detectors do not provide boundary contours, resulting in poor grasp orientation on randomly stacked items.
Conceptual Signal Flow
System Concept
Wrist RGB-D Camera → Semantic Segmentation Net → Grasp Pose Estimator → Gripper Actuation.
Architecture Modules
1Wrist-Mounted Compact Depth Sensor
2Edge Neural Tensor Processing Unit
36-DOF Grasp Pose Synthesis Module
4Adaptive 2-Finger Gripper
Hardware Categories
Compact RGB-D SensorParallel Jaw GripperEmbedded Edge ProcessorRobot Arm Joint Interfaces
Technologies & Software
PyTorchONNX RuntimeComputer VisionROS 2 MoveItGrasp PlanningFastSAM / YOLO-World Mask ModelsGrasp Synthesis Algorithms
Engineering Considerations & Edge Cases
- Cable flex fatigue on camera tether lines during full wrist rotation
- Inference latency budget within arm motion trajectory
Applications
Random bin-picking in factories
Recycling sorting stations
Agricultural harvest sorting
Engineering Inquiry & Collaboration
Interested in Developing this Architecture?
Connect directly with Tamizh Tech engineers in Coimbatore to discuss mechanical fabrication, sensor selection, firmware implementation, or turnkey system commissioning.