SUBMO

Despite significant advances in artificial intelligence to solve logic and language-based problems, robots still struggle with tasks involving physical interaction, which humans perform almost subconsciously. The paradox of human performance lies in our ability to excel at these tasks despite slower neural response times and greater system complexity when compared to robotics. This gap in understanding affects both artificial intelligence and neuroscience; it limits our ability to leverage prosthetics, exoskeletons, and stroke rehabilitation to develop effective solutions for activities of daily living (ADLs) -- tasks that inherently require physical interaction.

Approach: In this Ambizione grant, I propose to leverage mechanical physics to subtract physical interaction and peripheral biomechanics in order to uncover neural factors. The insights gained will inform generative models to advance both motor control and robotics. This will be achieved through three work packages (WP):

-WP1: Subtractive Modeling. An instrumented-actuated crank will be fabricated, novel system identification algorithms will be developed, human-subject crank turning experiments will be conducted, and subtractive methods will be employed to estimate the zero-force trajectory and mechanical impedance. Then, this information will be used to test hypotheses related to the theory of Elementary Dynamic Actions.

-WP2: Generative Modeling with Submovements. A submovement-based model predictive control strategy will be developed; it will be extended to investigate redundancy, and next, it will be adapted to manage variable impedance. Simulations will test hypotheses about the structure of the human controller by comparing the simulation results to data collected in WP1. These controllers will then be implemented on a physical dual-arm Franka system using the hardware developed in WP1. Finally, the complete controller will be applied to a more challenging task: planar bar manipulation, which will be detailed below.

-WP3: Generative Modeling with Central Pattern Generators (CPG). As a first pass, system identification will be performed on existing CPG-based locomotion models; next, a CPG-based model will be developed to simulate the crank-turning experiment in the medium speed case; it will be trained using reinforcement learning, and system identification methods will be performed on the model to estimate the zero-force trajectory and mechanical impedance. These estimates will be used to test hypotheses about CPG-based control.

The ideas herein develop descriptive models of human motor control that are directly applicable to prosthetics, exoskeletons, and stroke rehabilitation. Yet, these ideas have far broader implications. We are at a pivotal time in robotics, where advancements are happening rapidly, but significant limitations remain in managing physical contact. In the future, I believe that robotic systems featuring both mechanical and software-controlled variable impedance will become the standard. Human evolutionary biology has elegantly demonstrated a solution to this problem. However, this remains an unsolved research challenge. This Ambizione grant directly addresses this puzzle. Answering this question will not only help people with physical disabilities, it will also result in algorithms that manage physical interaction despite delays and kinematic complexity. The answer to this question could redefine how people view robotic systems - from rigid mechanical devices to those capable of seamlessly performing activities of daily living.