Orientation Estimation of Robotic Gripper to Grasp Activities of Daily Living (ADL) Object Using YOLOv5-OBB

Research Objectives: To predict ideal grasping angle of Activities of Daily Living (ADL) object, accomplished with object-detection model YOLOv5-OBB, which will be used by UFACTORY xArm 5-DoFs robotic arm fitted with gripper and camera.

Design: Experimental study

Setting: Evaluation of machine-learning model using digital images, then implementation of model in BioRobotics Lab at University of Wisconsin-Milwaukee

Participants: Healthy participants

Interventions: At this stage of project (within scope of this abstract), YOLOv5-OBB object-detection model is trained to detect ideal grasping angle to be adopted by robotic gripper. Model is implemented in physical system, and robotic gripper rotates according to model's output. After completion of this stage (beyond scope of this abstract), model is trained to localize object and estimate width. Finally, participants are seated at wheelchair where robot is mounted and equipped with object-detection model so that robot can locate, grasp, and pick up object that is situated on adjacent table.

Main Outcome Measures: Validation images:

1) Mean Angular Error - expresses average magnitude of error of predicted grasping angle
2) Precision
3) Recall
4) Mean Average Precision (mAP)

Physical system:

1) Mean Angular Error

Results: Validation images:

1) Mean Angular Error: 6.95 degrees - satisfactory for many grasping situations
2) Precision: 66% - good
3) Recall: 72.2% - good
4) mAP@0.5: 63.2% - good

Physical system:

1) Mean Angular Error: TBD with sufficient number of samples, but in two observations performed so far, one sample had 0 degrees error and another sample had approximately 5 degrees error.

Conclusions: Task of predicting ideal grasping angle is accomplished by training YOLOv5-OBB model and evaluating on validation images. Validation results, especially Mean Angular Error, are quite satisfactory. Efficacy of model in physical system has been observed; performance in physical system will soon be quantified using sufficient number of samples to calculate Mean Angular Error. Future work consists of localization and estimation of width of object before model is fully operational and implemented in system with participants.

Author(s) Disclosures: None