Vision-Guided MPC for Robotic Path Following Using Learned Memory-Augmented Model.

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Precise robot-environment interaction is crucial for contact-rich tasks like machining and polishing.
• Adapting to uncertain environments and dissimilar objects is vital for emerging fields like disassembly and recycling.

Purpose of the Study:

• To develop an end-to-end framework for trajectory-independent robotic path following in uncertain environments.
• To improve the accuracy and adaptability of robotic control for contact-rich tasks.

Main Methods:

• Combined model predictive control with image-based path planning and real-time visual feedback.
• Utilized a learned state-space dynamic model incorporating a differentiable neural computer (a type of memory augmented neural network).
• Introduced a novel application of memory augmented neural networks for modeling robot-environment dynamics during contact.

Main Results:

• Achieved a 21.0% reduction in Root Mean Square (RMS) error compared to a Long Short-Term Memory (LSTM) architecture.
• Demonstrated the framework's ability to generalize to previously unseen materials in simulations.
• Validated the framework's effectiveness in handling parametric uncertainties.

Conclusions:

• The proposed framework offers a significant advancement in robotic path following for contact-rich tasks.
• Differentiable neural computers show promise as a superior alternative to LSTMs for dynamic system modeling in robotics.
• The trajectory-independent approach enhances adaptability for real-world applications with unpredictable conditions.