UCLA

This collection of studies explores various facets of robotic intelligence, emphasizing the lifecycle of a robot from inception to advanced operation. Beginning with the design phase, the concept of ``physical intelligence" is demonstrated using the Non-Anthropomorphic Biped-Soleus (NABi-S) robot. The NABi-S robot, with its unique leg alignment and compliant soleus mechanism, demonstrates stability and agility while allowing for recovery from perturbations without complex control systems. Simplified kinematic analyses and open-loop control algorithms are all that are necessary for natural motion.

In cases where closed-loop control is necessary, intuitive design approaches are taken to hybridize the CPG controller with zero dynamic control methods in order to render marginally stable and unstable zero dynamics of a cart-pole system, stable. Without the use of rigorous control theory, empirical tuning approaches are all that is necessary to synchronize a CPG state to the cart's pivot state to robustly stabilize the system. We then extend closed-loop CPG control to a hybrid dynamical system, the Simplest Walker, and show that stability regions surrounding walking cycles can be extended. This leads to the discussion of the synergistic benefits of hybridizing a CPG controller with an HZD controller. Through analysis, it is found that robots like NABi-S that have a higher number of degrees of freedom (DOFs) and walk unconventionally can be controlled successfully with this algorithm.

In consideration of communication and cooperation, it is shown that the integration of Multi-Input Multi-Output (MIMO) network theory enhances spatial intelligence and awareness among UAV teams, allowing for high-level trajectory optimization in simultaneous data aggregation and communication tasks. By making small adjustments to pre-existing sensing tasks, this approach achieves significant improvements in network efficiency, demonstrating how ``spatially aware" modifications to data collection plans can lead to substantial gains in performance.

Lastly, for robotic perception, SLAM (Simultaneous Localization and Mapping) techniques are crucial. The Block Online Expectation Maximization (BOEM) SLAM algorithm presents a robust hybrid approach to visual-inertial navigation that combines filtering and optimization techniques. Unlike optimization-based SLAM methods, this method does not require the processing of an entire batch of data at once, yet it achieves greater accuracy than filtering methods alone. By efficiently fusing visual and inertial data, BOEM-SLAM supports accurate realtime localization while lowering hardware requirements, making it useful for the simplification of robotic systems by algorithmic means.

Together, these studies underscore the multifaceted intelligence required for modern robots, from foundational physical design to advanced spatial and visual capabilities.