A CyberOctopus that Learns, Evolves, and Adapts
Abstract
Our objective is to develop a computational analog to living octopuses, a CyberOctopus that adapts, learns, and evolves to novel tasks, situations, and environments. Our rationale is that this software suite will drive the convergence and synthesis of novel control theories and algorithms, simulation of soft-bodied creatures, and new insights into cephalopod neurodynamics enabled by unprecedented in-vivo sensing techniques. Overall, our efforts will provide a framework towards the integration of octopus biophysical principles into modernengineering design and control systems, and will form a solid foundation to enable further investigations beyond the time-frame of this MURI.Our technical approach relies on four interconnected thrusts: In RT#1 we will design a set of biologicals experiments to elucidate the neurodynamics principles underlying octopus~ abilities, and ontogeny of behavioral adaptation in cephalopods. RT#2 will bring together established and novel sensing technologies into a unique ~full view~ testing environment to simultaneously acquire in-vivo and ex-vivo data from freely moving octopuses. These sensing technologies include novel conformal and wireless electronics to be surgically implanted, machine vision systems that can detect behavioral primitives, and high resolution pressure sensors patterned in the octopus~ environment. In RT#3 we will develop a virtual environment in which dynamic models of octopus neurons and derived control abstractions (RT#4) can be tested in abiophysically accurate muscular system. To simulate the highly compliant octopus body we will use a novel approach based on Cosserat rods, and model its neural infrastructure using data from RT#1-2 relative to its different organizational levels. RT#4 will distill experiments and simulation results into theory, algorithms, and software to enable the CyberOctopus to learn complex sensorimotor tasks quickly from experience. Our preliminary hypothesis is that deep oscillators (directed acyclic graphs of dynamic neurons modeledas oscillators) integrate sensory information along the tentacles and leverage body mechanics to construct a probabilistic belief about the system state. This belief state is coded in the neurodynamics of the oscillator (neuron) population and not only provides the animal with a sense of the body (proprioception) and itsambient environment, but also informs the motor control strategies for locomotion, prey-handling, and problem solving at multiple temporal scales. Our work will result in new learning models fundamentally more capable of handling distributed dynamic systems characterized by large numbers of degrees of freedom. We have assembled a powerful team of biologists, engineers, and mathematicians with complementing expertise. Through the work performed in this MURI we will build a dialog between the broader research community, and in particular between researchers in cephalopod biology and biological neuronal networks, with researchers in machine learning, control, and robotics. Towards this, we will support summer fellowships, seminars, and workshops to facilitate wider engagement with other researchers. In addition, the data, software, models, and protocols developed in this MURI will be made openly available through a website we will maintain.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 23, 2019
- Source ID
- N000141912373
Entities
People
- Girish Chowdhary
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign