Curriculum, Not Touch, Drives Robotic Learning of Object Manipulation

Engineers show that robotic hands can learn complex object manipulation without touch—redefining the role of sensory feedback in machine and biological learning.

Research: Curriculum is more influential than haptic feedback when learning object manipulation. Image Credit: JQ SNAPS / Shutterstock

How does a robotic arm or a prosthetic hand learn a complex task like grasping and rotating a ball? The challenge for the human, prosthetic, or robotic hand has always been correctly learning to control the fingers to exert force on an object. The sensitive skin and nerve endings that cover our hands have been attributed to helping us understand and adapt our manipulation, so roboticists have insisted on incorporating sensors into robotic hands. But–given that you can still learn to handle objects with gloves on– something else must be at play. 

This mystery inspired researchers in the ValeroLab in the Viterbi School of Engineering to explore whether tactile sensation is always necessary for learning to control the fingers. The researchers Romina Mir, Ali Marjaninejad, Andrew Erwin, and Professor Francisco Valero-Cuevas, within the Alfred Mann Department of Biomedical Engineering, asked: how do the sensors that are part of the hand (its nature) interplay with how a hand is trained (nurtured) to learn complex tasks?

In a paper in the journal Science Advances, the team addresses the classic "nature versus nurture" question using computational modeling and machine learning. The paper "Curriculum Is More Influential Than Haptic Information During Reinforcement Learning of Object Manipulation Against Gravity" builds on the lab's previous work related to hand evolution and artificial intelligence. It demonstrates that the sequence of learning, also known as the "curriculum," is critical for learning to occur. In fact, the researchers note that if the curriculum takes place in a particular sequence, a simulated robotic hand can learn to manipulate with incomplete or even absent tactile sensation.

For this study, the team employed software to simulate a three-finger robotic hand to "provide a counter-example to the long-held notion that tactile sensation is always necessary," Valero-Cuevas says. The study also "emphasizes the importance of the sequence of rewards for training," commented Romina Mir, one of the two first authors and a PhD student in the ValeroLab.

"Reward guides development of the system," said corresponding author Francisco Valero-Cuevas, a professor in the Division of Biokinesiology and Physical Therapy at USC.

He added, "…just like biological systems are a product of their experience. This link between machine learning and biology is a powerful connection that may enable progress of artificial intelligence systems that can learn and adapt in the physical world."

In this collaboration between the Viterbi School of Engineering and the University of California, Santa Cruz (UCSC), doctoral students Parmita Ojaghi (UCSC) and Romina Mir (USC) co-led this work in collaboration with Prof. Michael Wehner (UCSC). Ali Marjaninejad and Andrew Erwin (USC) also contributed to this work.