Candidate: Brandon James DeHart
Title: Dynamic Balance and Gait Metrics for Robotic Bipeds
Date: May 29, 2019
Time: 10:00 AM
Place: EIT 3142
Supervisor(s): Kulic, Dana - Gorbet, Robert B. (Knowledge Integration)
Abstract:
For legged robots to be useful in the real world, they must be able to balance and walk reliably. Both of these abilities improve when a system is more effective at moving itself around relative to its contacts (i.e., its feet). Achieving this type of movement depends both on the controller used to perform the motion and the physical properties of the system. Although much work has been done on the development of dynamic controllers for balance and gait, only limited research exists on how to quantify a system's physical balance capabilities or how to modify the system to improve those capabilities.
From the control perspective, there are three strategies for maintaining balance in bipeds: flexing, leaning, and stepping. Both stepping and leaning strategies typically depend on balance points (critical points used for maintaining or regaining balance) to determine whether or not a step is needed, and if so, where to step. Although several balance point estimators exist, the majority of these methods make undesirable assumptions (e.g., ignoring the impact dynamics, assuming massless legs, planar motion, etc.).
From the physical design perspective, one promising approach for analyzing system performance is a set of dynamic ratios called velocity and momentum gains, which are dependent only on the (scale-invariant) dynamic parameters and instantaneous configuration of a system, enabling entire classes of mechanisms to be analyzed at the same time.
This thesis makes the following contributions towards improving biped balancing capabilities: First, a dynamic bipedal controller is proposed which uses a 3D balance point estimator both to respond to disturbances and produce reliable stepping. Second, a novel balance point estimator is proposed that facilitates stepping while combining and expanding the features of existing 2D and 3D estimators to produce a generalized 3D formulation.
Third, the momentum gain formulation is extended to 2D and 3D systems, then both gains are compared to centroidal momentum via a spatial formulation and incorporated into a generalized gain definition. Finally, the gains are used as a metric in an optimization framework to design parameterized balancing mechanisms within a given configuration space. Effectively, this enables an optimization of how well a system could balance without the need to pre-specify or co-generate controllers and/or trajectories.
To validate the control contributions, simulated bipeds are subjected to external disturbances while standing still and walking. For the gain contributions, the framework is used to compare gain-optimized mechanisms to those based on the cost of transport metric.
Through the combination of gain-based physical design optimization and the use of predictive, real-time balance point estimators within dynamic controllers, bipeds and other legged systems will soon be able to achieve reliable balance and gait in the real world.