Simulation-Aided Handover Prediction From Video Using Recurrent Image-to-Motion Networks.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks offer advancements in visuomotor robot learning.
• Predicting future poses is crucial for human-robot and robot-robot collaboration dynamics and safety.

Purpose of the Study:

• To propose a novel recurrent neural architecture for visuomotor robot learning.
• To enable accurate prediction of robot trajectories from motion videos for cooperative tasks.

Main Methods:

• Developed a recurrent neural architecture to transform variable-length motion videos into robot trajectory parameters.
• Utilized a simulation environment to expand the training database and enhance network generalization.
• Trained models on both synthetic and real-world data.

Main Results:

• The proposed architecture accurately predicts handover trajectories, even with limited input frames.
• Models trained on combined synthetic and real data outperformed those trained on single data sources.
• Successfully executed handover tasks with uncalibrated robots using computed trajectories.

Conclusions:

• The novel recurrent neural architecture effectively predicts robot trajectories for cooperative tasks.
• Combining synthetic and real data improves model performance and generalization.
• The approach facilitates cooperative tasks like object handover with uncalibrated robots.