Reproduce “Force Generative Imitation Learning: Bridging Position Trajectory and Force Commands through Control Technique” on OpenResearch

About this paper

Hiroshi Sato, Sho Sakaino, Toshiaki Tsuji

Researchers from the University of Tsukuba and Saitama University developed a hierarchical framework for robots to learn contact-rich tasks, generating appropriate force commands from desired position trajectories. This "Force Generative Imitation Learning" method, by integrating a memoryless lower layer with PID control, achieved superior tracking accuracy and generalization in character writing tasks compared to conventional methods.

Key insights

• Directly integrating external feedback control with memory-based neural networks (like LSTMs) can lead to instability and inhibit performance.
• Employing a memoryless neural network (MLP) as a lower-layer controller enables stable and effective integration of classical feedback control.
• Time-scale separation, with memory in the slower upper layer and a memoryless, fast lower layer, is a viable design principle for stable learning-based force generation.

Summary by alphaXiv.