Motion Prediction in Collision Avoidance Manual Reaching Tasks Considering Cognitive Perceived Risk

Human motion and coordination involve the complex interplay of cognitive perception and biomechanical capabilities. This connection allows individuals to perform tasks such as reaching, grasping, and navigating through obstacle-filled environments with remarkable efficiency. This dissertation investigates human collision avoidance strategies during manual reaching tasks by focusing on the minimum clearance distance between the upper extremity and three-dimensional obstacles of varying characteristics. By integrating biomechanical and cognitive factors, this research advances the understanding of how cognitive perceived risk influences human motion in obstacle-populated environments. Experimental data were collected from fifteen participants as they performed natural reaching motions around three distinct obstacles using an IMU-based motion capture system. These obstacles varied in shape, size, orientation, and fragility, prompting participants to adjust their motion strategies based on risk perception. This work proposes and tests multiple loss functions to quantify the cost of collision with an obstacle, including a risk factor loss function tailored to each obstacle and parameterized using experimental data. Other loss functions, such as the log-cosh, tilted, and weighted absolute error loss functions, are also evaluated, offering a broader perspective on how different mathematical formulations influence motion prediction. The study incorporates a Bayesian Decision Theory framework to model perceived collision risk, an approach that quantifies the cognitive biases involved in human decision-making under uncertainty. The optimization-based motion prediction algorithm was enhanced by integrating the proposed loss functions as constraints, allowing for the prediction of joint angle profiles and minimum clearance distances. A sensitivity analysis for the tilted and weighted loss functions demonstrated that the parameter 𝜏 plays a crucial role in improving joint angle profiles’ predictions. The results show that the prediction accuracy for shoulder and elbow joint angles was higher than for wrist motions, as evidenced by lower root mean square error values in the joint angle profile comparison with experimental data. Among the loss functions, the risk factor loss function proved most effective for minimum clearance distance prediction, likely due to its obstacle-specific parameterization. However, each loss function exhibited distinct advantages and disadvantages, demonstrating the importance of selecting a loss function based on the specific requirements of the simulation task. The results further demonstrate that obstacle characteristics significantly affect the minimum clearance distance, with wider and more fragile obstacles resulting in greater clearance. No significant differences were found in motion duration or maximum velocity, suggesting that these factors may not directly correlate with perceived risk in static environments. Overall, including perceived risk through the proposed loss functions led to more accurate predictions than traditional non-risk-based simulations, highlighting the importance of including cognitive factors and biomechanical constraints. This research contributes to human motion modeling by combining biomechanical efficiency with cognitive factors, such as perceived risk, into a unified prediction algorithm. The findings highlight the need to incorporate cognitive models of risk perception into optimization-based motion prediction algorithms for applications such as human-robot interaction, ergonomics, and digital human modeling.