PERSONALIZED, VISION-BASED HUMAN-LIKE ROBOTIC ARM CONTROL ON REAL-TIME EMBEDDED SYSTEMS

Sammanfattning

This work investigates how human-like and personalized robotic arm motion can be generated and executed under real-time embedded constraints using Dynamical Movement Primitives (DMPs). The area is of increasing interest due to the growing use of robotic systems in assistive robotics, rehabilitation, collaborative robotics, and other human-centered applications where natural and adaptive motion can improve usability, safety, and user acceptance. Achieving human-like robotic motion while maintaining predictable real-time behavior is challenging, particularly on embedded systems with limited computational resources. The work focuses on three main objectives: evaluating motion quality relative to human demonstrations, investigating the effect of personalization on perceived movement similarity and human-likeness, and analyzing the timing characteristics of the complete motion generation pipeline in a real-time setting. Human motion demonstrations were recorded using an OAK-D Lite camera and pose estimation models during a cup-moving task. Joint trajectories were extracted from the recorded keypoints and used to train DMP-based motion models. To incorporate personalization, subject-specific curvature-based coupling terms were integrated into the DMP formulation to adapt the generated trajectories toward individual movement tendencies. The proposed system was evaluated through three experiments. The first experiment quantitatively compared generated motion against human demonstrations using Log Dimensionless Jerk (LDLJ), Spectral Arc Length (SPARC), and Root Mean Square Error (RMSE). The second experiment investigated perceived personalization and human-likeness through a qualitative user study. The third experiment evaluated the timing behavior of the embedded motion generation pipeline by measuring end-to-end latency, stage-wise execution times, jitter, and tracking performance. The results demonstrated smooth, repeatable, and low-latency robotic motion generation suitable for real-time embedded robotic control.