Estimation of task-specific manipulability scores for a robotic manipulator in vacuuming scenarios

Abstract

Reliable vacuum-cleaning trajectories require selecting joint configurations from which a robotic arm can continue motion in task-relevant directions without encountering self-collision, joint-limit violations, or singularities. Classical manipulability measures describe dexterity using the Jacobian but do not account for the geometric constraints and directional motion patterns typical in floor-cleaning tasks. This work presents a data-driven method for estimating a task-specific manipulability score that reflects how easily a seven degrees-of-freedom robotic arm equipped with a floor-contact vacuum tool can continue motion in six primitive directions (forward, backward, left, right, clockwise rotation, counterclockwise rotation). A dataset of one hundred and seventy-six thousand and twenty valid joint configurations was generated by sweeping a four-dimensional grid of feasible end-effector poses, computing up to three inverse-kinematics solutions per pose, and simulating incremental movements with collision checking. Each configuration was assigned a directional score based on inverse-kinematics reachability, collision outcomes, and distance-dependent penalties.

A fully connected neural network was trained to regress six scores from the seven joint angles. The model achieved a denormalized Mean Absolute Error of approximately two point zero for translational directions and one point fifty-five to one point sixty for rotational directions (approximately eight percent of the full score range), while enabling extremely fast inference–around ninety-five thousand eight hundred and eighty evaluations per second on a consumer GPU.

By shifting the computationally expensive stage of collision checking and inverse-kinematics sampling to offline preprocessing, the method provides a lightweight surrogate for online motion planning. The learned score can help planners avoid unfavorable configurations and maintain consistent vacuuming trajectories. Limitations include the absence of environment-aware terms, stochastic inverse-kinematics sampling, and hand-tuned scoring parameters. Future work will focus on integrating obstacle information, improving label generation, and embedding the score into trajectory optimization frameworks for more robust real-world operation.