Science of Embodied Innovation, Learning and Control
Abstract
A robots mobility depends on its capacity to move energy from a store in its mass center along the right degrees of freedom at the right time by actuating appendages toward the periphery where it meets its environment. Because there is a premium on getting this work done quickly, power (the rate at which actuators can move Joules) is a first scarce resource. The information required to direct these outward flows appropriately must also be generated from some prior memory combined with feedback decisions made using real time streams. Moreover, since the purposes of mobility are inevitably linked to the robotÕs knowledge about the environment as well as the task, its ability to bring information from the periphery inward to the core at adequate rates inevitable presents a challenge simultaneous with and dual to its management of outward power flows. The design of morphology Ð the nature of limbs and body and their endowment with actuation and perceptual resources Ð to promote effective interaction between energy and information streams over contrasting scales of length and time represents the first focus of our project. The discovery of how to evolve, use and revise this endowment to achieve goal-directed mobility and create new solutions to sensorimotor limitations and challenges represents the second focus. We propose a fundamental inquiry into the design and use of morphology as the seat of embodied agency and creativity over scales of length and tome covering multiple orders of magnitude. Our research aims to discover: (RCA1) universal physical constraints on rates of energy-information exchange that can expose design implications at key points of interest along the vast spatiotemporal scales; (RCA2) the endowment of animal body plans and materials with actuation and perceptual resources as the foundation of learning to engage the physical environment at evolutionary, developmental, and behavioral timescales in the face of these fundamental constraints; and (RCA3) the implications of such organismal innovation within these physical limits for robot design and deployment. Impact on DoD: RCA1 aims for new fundamental insights into limits on the mathematical structures and information thermodynamic mechanisms underlying the algorithms and supporting physical design of purposive mechanical work. The identification of a new universal limit bearing on sensorimotor systems opens the way for a host of assays that the DoD service labsÕ bench scientists and engineers can then develop for measuring the efficacy of existing proposed designs. RCA2 take a complementary step, distinct from the contemporary AI focus on deep learning, to gain new inspiration from the mammalian hippocampus. Formalizing the neuromechanical basis of animalsÕ abilities to innovate promises DoD new, computationally tractable representations for higher order learning that could be joined to rapidly advancing reinforcement learning technologies. Finally, because models of unstructured (e.g., gravel- strewn, broken-sloped, icy or leaf-littered) substrates are so incomplete, we require physical experiments with robots in natural terrain to test our hypotheses. To benefit from animal designs and pursue fundamental tradeoffs demands far more controlled and rational materials use that can be found in commercial robot designs. Thus, RCA3 focuses on novel materials for distributed sensorimotor structures in robotics, directly impacting DoD capabilities through the development of new robots with unparalleled mechanical competence and autonomy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810327
Entities
People
- Daniel E. Koditschek
Organizations
- Army Contracting Command
- United States Army
- University of Pennsylvania