This EArly-concept Grant for Exploratory Research (EAGER) collaborative project will explore a novel interdisciplinary approach to motion planning and control for robust and capable legged machines, inspired by the way that humans navigate challenging terrain, and implemented using computational methods originally developed for space missions. Whether for agriculture, construction, or disaster response, the mobility afforded by legs offers promise for future robots that can go where people go, either in their place or by their side. Yet, for these future robots to succeed in real-world environments they must -- as humans do -- plan and execute purposeful movements, to respond to unexpected obstacles as they arise. Human capabilities in this regard vastly exceed those of modern robots. When humans move through the world, they appear to adjust the level of complexity that they include in their planning to the immediacy of the need. That is, movements that must be executed within the next few moments are visualized in detail, while those that will not occur for some time are abstracted more coarsely. The control scheme created in this project will apply a similar approach in legged robots, to achieve safe, precise movements guided by a long-range strategy. The results will advance the national prosperity and welfare, by enabling legged machines that can make the rapid decisions necessary to keep their balance and avoid falls, improving robustness for practical deployment as first responders, home health aides, explorers, or co-workers.
The project will lay the foundation for a new paradigm of robot control that makes use of a rigorous methodology for optimal control over hierarchical abstractions with a novel computational solution framework. Optimal control over hierarchical abstractions will provide a new tool for control designers, allowing them to strategically enforce consistency between coarse-grained long-term plans and fine-grained near-term control. In legged robots, the absence of a rigorous framework for managing such a challenge has 1) prevented practical hardware implementation of full-model trajectory optimization and 2) limited the robustness of control based on simple models. In this work, these separate approaches will be unified and their combined benefits captured. The envisioned solver uses a new multiple-shooting formulation to reduce problem sensitivity and a new quasi-Newton approximation to reduce runtime. Fundamental efforts as part of the EAGER effort will consider control synthesis for simplified 2D models of quadruped bounding and biped running. The generality of envisioned abstractions for these cases will be studied. Necessary control rates will be assessed in simulation, and a breakdown of the computational requirements for different algorithm components will be determined. This data will be critical to identify further advances to the approach that will be necessary for its future use as an online control method to stabilize locomotion in 3D robots.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
