S&AS: FND: COLLAB: Probabilistic Underactuated Motion Adaptation

University Of Rochester, Rochester NY

Investigators

Abstract

Unconventional, underactuated robots, such as humanoids or legged platforms more broadly, offer the potential to move through and perform work in constrained, three-dimensional environments that are currently inaccessible to existing autonomous agents. However, this potential has been largely unrealized because it is difficult to reliably adapt the behaviors of these platforms to account for the changing and uncertain task and environmental conditions in the "real world." Although many of the fundamental principles that govern contemporary task and motion planning techniques are applicable across different platforms, the practical implementation of these principles has been largely platform specific. In contrast, this project will adopt a probabilistic planning framework which learns common structure for the motion patterns of different platforms performing related tasks, then uses this structure to generate generalized, inherently platform independent, motion primitives. At runtime, the primitives will be grounded and adapted where necessary to specific robot models given local task and environmental conditions. The primary benefit of this project will be an increase in the utility of autonomous platforms for tasks such as urban search and rescue, industrial inspection, and planetary exploration. The analytical techniques that will be developed will have further impacts on locomotion science and learning-based approaches to motion coordination. The PIs will additionally be involved with K-12 outreach involving robot demonstrations at FIRST Robotics Competitions and the Rochester Museum and Science Center. This project will specifically address fundamental limitations in the tractability of real-time task and motion planning for underactuated robots over diverse objectives and distributions of environmental conditions. Probabilistic models will be developed to efficiently reason over and adapt the nominal behaviors of different highly-articulated, underactuated robots. The behavioral inference will make it possible to 1) select appropriate pre-existing behaviors (developed over the course of the project) where relevant, 2) use novel combinations of nominal behaviors to form compound, task-specific behaviors, and 3) leverage similar, but not necessarily the same, kinematic structure across heterogeneous platforms to transfer behaviors between them. To ensure the success of the practical, online implementation of the developed models, the PIs will develop algorithms that combine probabilistic inference, nonlinear dimensionality reduction, and dynamic movement primitives to produce a novel combination of efficient motion generation and robust online adaptation. In addition to varying task and environmental conditions, the adaptability of the probabilistic models to changes in the internal kinematics and dynamics of robot platforms, such as those that would arise from degraded motor performance or structural failures of joints or entire limbs, will also be explored. The models will be trained and validated using a combination of simulation and experimental results on two physical platforms: the Carnegie Mellon Hexapod and the Robotis OP2. Furthermore, the PIs will develop software tools and release open-source products related to generalizable probabilistic models for motion adaptation of underactuated systems.