Introduction
For robots to function effectively as human assistants, they must transition from reactive behaviors to proactive engagement. In the context of object handovers, a robot must anticipate a human's intent early in the movement to optimize its gripper's position and orientation. This task is complicated by the non-linear nature of human motion, the necessity of modeling coupled 6D geometry (translation and rotation), and the need for low-latency classification. This paper introduces a framework that utilizes Gaussian Processes (GPs) and dual quaternions to achieve high-precision trajectory prediction and early intent recognition.
Methodology
The proposed system employs a hierarchical approach to model and classify human motion.
Dual Quaternions for 6D Motion
Traditional methods often treat 3D position and 3D orientation as independent variables, which can lead to mathematical singularities like gimbal lock and ignore physical correlations. The authors utilize dual quaternions, a unified 8-element vector representation that captures both rotation and translation. This ensures that the relationship between hand movement and orientation is preserved throughout the modeling process.
Gaussian Processes on Manifolds
The researchers extended standard Gaussian Processes to operate within the space of dual quaternions. By defining a custom distance metric ($d_{mag}$) and a specialized squared exponential kernel, the GP can effectively handle the non-Euclidean geometry of rotations on a sphere.
Stochastic Trajectory Classification
To manage various handover styles, trajectories are clustered and compared against live data using a normalized inverted Mahalanobis distance. The system employs a set of decision rules—including probability thresholds and moving window integrals—to "prune" unlikely handover configurations in real-time.
Experimental Evaluation
The framework was validated using a dataset of human-human object handovers. Participants performed handovers of a cylinder under 10 distinct conditions, varying grasp types (bottom, top, above, reversed) and target orientations (right, down, left, up). Data was captured using a high-frequency magnetic motion tracking system.
Key Results
The system demonstrated exceptional performance across several metrics:
- Classification Accuracy: 100% accuracy was achieved across all 50 test trajectories.
- Early Intent Recognition: On average, the system nominated the correct handover configuration at just 43.4% of the total trajectory length, allowing significant time for robotic response.
- Prediction Precision: Post-nomination, the GP predicted the remaining trajectory with a Root Mean Squared Error (RMSE) of 0.0029, aligning with natural human movement variability.
Conclusion and Future Work
By combining unified 6D modeling with stochastic classification, this framework enables robots to anticipate human actions with high confidence and speed. While the current model relies solely on kinematics, future iterations could incorporate visual context—such as object shape or the receiver's hand position—to further accelerate intent recognition. Additionally, scaling the GP training process through sparse methods will be essential for broader real-world applications.
Summary updated