Quadruped robots are versatile locomotion platforms capable of a variety of manipulation tasks. For each operating environment, the quadruped's choice of kinematic design, actuators, and end-effectors requires tuning and optimization. This research introduces geometrically enhanced designs for limb mechanisms and magnetic variable stiffness end-effectors for a ship-board quadruped platform, SORREL.

The choice of kinematics in quadrupeds is highly dependent on the gait and certain leg designs might be more prone to waste power. Traditional robots, design their kinematics based on the actuator being a single unit consisting of a motor and a gearbox, overlooking the effects of their relative positions on the limb kinematics. A new method was developed for designing the leg kinematics by decoupling the design of the gearbox and the motor. This method gives more freedom in the design of the kinematics and allows for a wider range of configurations best fit for the required task. The quadruped leg design presented in this work has both actuators collocated at the hip in a parallel configuration; however, the gearbox of the knee motor is relocated from the hip to the knee joint. This seemingly simple modification significantly reduces the mechanical antagonism and motor losses and increases the stiffness of the leg. Quadrupeds perform manipulation and grasping through task-specific actuators. These actuators need to interact with the environment safely while maintaining accurate force control capabilities. More specifically, SORELL needs to be equipped with gripper modules to deal with varying tasks such as typing on keyboards, precision manipulation of electronics, and general pick and place. Most grippers used in manipulators are based on stiff positioned-controlled actuators which provide accuracy but require bulky and expensive Force/Torque (FT) sensors for force control. Additionally, traditional grippers struggle with delicate handling tasks and pose safety risks due to their inherent high stiffness. By employing compliant actuators, namely nonlinear series elastic and variable stiffness actuators, robotic platforms can circumvent such limitations while maintaining the accuracy and stability of the end-effector.

As such, a family of actuators based on magnetic and electromagnetic forces was developed and tested to improve the safety of manipulation and grant accurate force control. The actuators presented use series elastic (SEA) and variable stiffness (VSA) concepts. Two actuators use passive nonlinear springs with constant stiffness profiles based on permanent magnet arrangements. These passive nonlinear SEAs, designed for lightweight and low-cost grippers, can achieve both stiff and soft behavior and have a higher capacity for energy storage than linear springs. With accurate spring modeling, these nonlinear SEAs can change the force control problem into a position control problem avoiding the use of FT sensors. Since passive SEAs use springs with fixed stiffness curves, their stiffness cannot be controlled independently. Moreover, changing the stiffness online is desirable in some grasping tasks that require adjustment in grip strength and compliance. Hence, two actuators that have active stiffness variation (VSA) capabilities through mechanical and electromagnetic means were developed. These actuators, while more complex and costly, give a wider range of stiffness and the ability to tune their stiffness curves. These actuators are studied and categorized for the intended manipulation tasks encompassing those required on the SORELL platform. Through these designs, multiple unique features were studied and observed such as water-cooled electromagnetic coils, eddy current damping for shock absorption, and the use of negative stiffness for fast explosive motion of gripper fingers.

The field of legged robotics has made significant advancements and shown potential practicality in various applications. Although these robots are becoming more popular, they are still not widely used due to the inherent danger when malfunctioning as well as their high cost. BALLU, Buoyancy Assisted Lightweight Legged Unit, is a robot that never falls down due to the buoyancy provided by a set of helium balloons attached to its lightweight body. This platform solves many issues that hinder current robots from operating close to humans while also providing affordability. However, the advantages gained also lead to the platform's distinct difficulties caused by severe underactuation and nonlinearities due to external forces such as buoyancy and drag. This dissertation presents a motion planning approach using data-driven techniques motivated by these challenges and its application to BALLU. The paper describes the concept of the platform, the hardware design of different generations of BALLUs, the software architecture, the nonconventional characteristics of BALLU as a legged robot, and an analysis of its unique behavior. Based on the analysis, a data-driven approach is proposed to achieve non-teleoperated walking: a statistical process is proposed to form low-dimensional state vectors from the simulation data, and a deep neural network-based controller is trained. The controller is tested on both simulation and real-world hardware. Its performance is assessed by observing the robot's limit cycles and trajectories in Cartesian coordinates. The controller generates periodic walking sequences in simulation as well as on the real-world robot, even without additional transfer learning. It is also shown that the controller can deal with unseen conditions during the training phase. The resulting behavior not only shows the robustness of the controller but also implies that the proposed statistical process effectively extracts a state vector that is low-dimensional yet contains the essential information of the high-dimensional dynamics of BALLU's walking.

For the past decades, humanoid robots have demonstrated increasingly impressive capabilities. However, the limitations of current generation of robots became painfully clear during the DARPA Robotics Challenge. They struggled with even basic locomotion under real-world conditions, a far cry from what was expected from the public. This study attempts to design a brand-new humanoid platform from ground up. Custom-designed proprioceptive actuators paired with liquid cooling are designed to maximize the torque density of the robot. The kinematics chain and structural parts of the robot are optimized for fast walking and running. Purpose-built sensors and electronics aim to provide robust feedback under impact load. Combined with new control architecture, the robot achieved untethered running as well as walking at over 2m/s. We demonstrated highly dynamic behaviors that are beyond the state-of-art and pushed locomotion capabilities one step closer to their human counterpart.

Although robotic locomotion and manipulation have shown some remarkable progress in the real world, the current locomotion and manipulation algorithms are inefficient in performance. They often only work for relatively simple tasks such as walking and running for locomotion and pick-and-place in structured environments (e.g., factory) for manipulation. In contrast, humans can perform quite dexterous tasks through contact as contacts provide additional dexterity to interact with environments. Hence, understanding the underlying contact mechanics plays a key role in designing contact-aware planners, controllers, and estimators for locomotion and manipulation.

However, design for planners, controllers, and estimators is extremely challenging. First, the number of contact states such as making and breaking contact with environments increases dramatically as the number of contacts increases. Thus, the underlying contact dynamics become large-scale non-smooth dynamics. As a result, optimization solvers have difficulties converging due to the non-convexity of the optimization problem.

Second, it is desirable that a robot should be able to interact in unknown environments during operation, leading to generalizable locomotion and manipulation. However, robust planning with frictional interaction with uncertain physical properties is very tough as the robot might cause undesired unexpected contact events. As a result, a robot might not be able to complete its desired task.

Third, once uncertainty is quite large, it is indispensable for closed-loop controllers to stabilize locomotion and manipulation. However, the design of manipulation is quite challenging as most manipulation systems are underactuated and unobservable with potential changes in contact states and modes.

In this dissertation, we present a methodology for contact-rich locomotion and planning using trajectory optimization. We first show that the planner using graph-search planners with trajectory optimization can be beneficial for decreasing the computation complexity. Second, we describe our contact-implicit trajectory optimization for planning of multi-limbed systems for running and climbing. We use decomposition-based optimization techniques to efficiently design a trajectory for a robot subject to various complicated contact constraints such as mixed-integer constraints. Then, we present our robust and stochastic trajectory optimization algorithms for multi-contact systems. We show that our chance-constrained optimization is applicable for planning multi-limbed robots. We also propose covariance steering algorithm for contact-rich systems using a particle filter to approximate a distribution of underlying contact dynamics. Our covariance steering is able to regulate robots' states and contact states simultaneously with probabilistic guarantees. Furthermore, utilizing the underlying structure of contact-rich manipulation, we present robust bilevel trajectory optimization for pivoting manipulation under uncertain physical parameters such as friction coefficients. Our proposed framework is able to design optimal control sequences while improving the worst-case stability margin along the manipulation. Finally, we present our closed-loop controller framework for tool manipulation using visuo-tactile feedback. Our approach enables the robot to achieve tool manipulation under unexpected contact events in closed-loop control fashion with no visual feedback for partially unknown objects.

The perspectives gained from this dissertation provide better insight into developing a contact-rich planning, estimation, and control framework for dexterous locomotion and manipulation in highly unstructured environments.

The properties of actuation is absolutely the basic and foundation among all factors that determine the overall performance of a system, and this has been true for all living creatures, and other machines as well. Ever since the first industry robot being invented half a century ago, the robotics field have been developed rapidly, and all sorts of robots are playing more and more important roles in all human production and daily life activities.

The main requirements for robots have been focused on rigidity, high torque and precision motion until recently, as traditionally the major application of robots are along factory production lines, with their working environment well known and controlled, performing jobs that are highly repetitive. To ensure the consistency of production, rigid robots that provides precise motion with high payload capacity are preferred. Thus, position controlled high torque servos have been dominating the industry. These actuators are also combined with force/torque sensors to enable force control capability on robotic systems, but due to various limitations of such setup, it is difficult to realize dynamic motions where impulsive loading is frequently recurrent. During the recent decade, research efforts in the robotics field have been made in the development of Series Elastic Actuators (SEA), seeking for alternative actuation solutions to meet the requirements for dynamic robotic systems. Although there are a hand full of successful cases, the complicity and difficulty down the route is quite obvious. Besides, the respond bandwidth of any SEA is strictly limited by the property of its elastic component, resulting in customization that is very application specific.

The recent rapid development of Proprioceptive Actuators, also know as Quasi-Direct Drive Actuators (or QDD for abbreviation), provides a simple actuation solution that not only features torque control capability, but also provides a very wide response bandwidth as well as inherent compliance. Thus its application on various robotic systems is definitely intriguing and worth of exploration.

The work presented in this dissertation involves the development and discussion of the application of Proprioceptive Actuators on several different dynamic robotic systems, including a three-finger robotic hand called DAnTE, a non-anthropomorphic bipedal robot called NABi-2 and a quadruped robot with very unique kinematics configuration, called ALPHRED-2. Underactuated anthropomorphic fingers are developed and applied on DAnTE, and a potential energy flow theory is developed to assist the development of the mechanical intelligence in these underactuated fingers to realize gesture dexterity. Along with the application of Proprioceptive Actuators on legged robots, an Impact Transfer Factor theory is proposed to quantify the response of a robotic limb to a impact in terms of impact sensing from a whole-system point of view. As outlook of the presented work, an upgraded version of DAnTE as well as an open source kid-size humanoid robot platform are introduced. Both systems are driven by well-packaged Proprioceptive Actuator modules and are expected to possess dexterous force control and high dynamic performance capabilities, and both of them are designed to serve as reliable robot platforms to carry further research in related topics.

Series Elastic Actuators (SEA) have been in development for multiple decades. In spite of this, few design guidelines exist and stiffness selection for the compliant element still remains a trial-and-error process. In this thesis, we experimentally validated the unlumped model first proposed by Orekhov for Rotary SEA(RSEA) and outlined a design methodology for selecting the spring stiffness based on the open loop force control bandwidth of unlumped model for series elastic actuators. We modified the unlumped model to apply to Rotary SEAs. Through experimental system identification, we demonstrated that our new unlumped model for RSEA is a valid model of actuator dynamics. Additionally, we recommended design guidelines for RSEA to achieve desired force control bandwidth based on the pure torque source assumption. An example of the design process was given and actuator performance was verified through dynamic simulations in ADAMS.

The control of locomotion on legged robots traditionally involves a robot that takes a standard legged form, such as the anthropomorphic humanoid, the dog-like quadruped, or the bird-like biped. Additionally, these systems will often be actuated with position-controlled servos or series-elastic actuators that are connected through rigid links. This work investigates the control implementation of dynamic, force-controlled locomotion on a family of legged systems that significantly deviate from these classic paradigms by incorporating modern, state-of-the-art proprioceptive actuators on uniquely configured compliant legs that do not closely resemble those found in nature. The results of this work can be used to better inform how to implement controllers on legged systems without stiff, position-controlled actuators, and also provide insight on how intelligently designed mechanical features can potentially simplify the control of complex, nonlinear dynamical systems like legged robots. To this end, this work presents the approach to control for a family of non-anthropomorphic bipedal robotic systems which are developed both in simulation and with physical hardware. The first is the Non-Anthropomorphic Biped, Version 1 (NABi-1) that features position-controlled joints along with a compliant foot element on a minimally actuated leg, and is controlled using simple open-loop trajectories based on the Zero Moment Point. The second system is the second version of the non-anthropomorphic biped (NABi-2) which utilizes the proprioceptive Back-drivable Electromagnetic Actuator for Robotics (BEAR) modules for actuation and fully realizes feedback-based force controlled locomotion. These systems are used to highlight both the strengths and weaknesses of utilizing proprioceptive actuation in systems, and suggest the tradeoffs that are made when using force control for dynamic locomotion. These systems also present case studies for different approaches to system design when it comes to bipedal legged robots.

Delivering packages from a warehouse to the customer's front door consists of a diverse set of sub-tasks, from packing packages into a truck to organizing them based on destination and even carrying them to a doorstep. Such a variety of behaviors can often be difficult for a single robot to achieve efficiently. Numerous research papers have proposed splitting the tasks amongst specialized robots with some dedicated to robotic arm manipulation and others to wheeling packages to their destination. Recently, companies have even started to pivot to the full generalized solution of humanoid robots. Although some success has been seen with both solutions, these technologies occupy large amounts of space, often have large costs, and have significant limitations that prevent parallelization and scalability. As a result, this reduces the amount and efficiency of overall package delivery capacity.

To resolve these bottlenecks, we introduced a new robotic platformed called LIMMS, or Latching Intelligent Modular Mobility System. LIMMS is a symmetric 6 degree of freedom (DOF) modular manipulator robot with a latching mechanism and wheels at both ends. In one configuration, LIMMS can use its end effector to latch itself to designated anchor points to behave like a traditional manipulator and move boxes around. Meanwhile, by placing anchor points on the box, multiple LIMMS can attach themselves to a box to act as legs in which the box is the body to facilitate transportation across larger distances including outside of the truck. In this way, all package delivery sub-tasks are covered by the different modes LIMMS can be in. The only limiting factor to the functionality of LIMMS is the anchor points. These anchor points, sometimes referred to as latching patterns or patterns, can be as close as a few inches or even tangent to each other given the box is sturdy enough. Many more LIMMS can work together to lift heavier boxes than any humanoid robotic system could. Not only has its physical constraints been reduced, but its spatial constraints as well. When parallelizing a task, having more robots can increase efficiency. However, too many robots working together can cause congestion leading to an inefficient system. With the latching mechanism, LIMMS can use any surface with latching patterns, e.g., walls, ceilings, boxes. This increases the overall operational surface area reducing the chance for gridlocks. With all of these features, LIMMS can be used to solve a large variety of problems more efficiently that no other system can due to its design.

This manuscript aims to detail the development and research findings of LIMMS. In particular there are five major contributions. The first being the introduction of LIMMS as a platform and the prototypes designed and built. Secondly, the latches on LIMMS are designed in such a way that it admits a large region for which latching is theoretically guaranteed. This theoretical boundary exceeds existing self-aligning mechanisms. The third contribution is the use of Jacobian fields to determine trajectories in which the load at the end effector is balanced through the structure of the robots, such that the effective torque required is much lower. Fourthly, we formulate an optimization problem to generate trajectories for LIMMS to use to deliver packages. Lastly, we demonstrate LIMMS as a concept with four LIMMS connecting to a box to form a quadruped to deliver itself and then return to the truck.

This thesis proposes a novel mechanism and its control strategy for generating a controllable resistance force for exercise and rehabilitation. Instead of measuring force with a load cell, the idea of a series elastic actuator is applied to perform a feedback force control by calculating force from spring deflection. By introducing series elasticity to an exercise machine, the proposed mechanism is able to achieve the active resistance force control under safe operations and in a cost effective way compared with existing systems. A prototype hardware was manufactured and an impedance control algorithm was developed in order to simulate the effects of a virtual damper, spring and constant force. The hardware design and control algorithm were examined through dynamic simulation and a prototype hardware was used to further verify the ideas.

The goal of this work is to formulate algorithms that can address three key ingredients I believe are necessary towards making robots autonomous and smart: (1) The robot needs to be able to react in an energy-efficient manner to outside disturbances; (2) The robot needs to understand its location of its surroundings and evaluate the uncertainty of its location to optimally and safely achieve some goal state; (3) The robot should continuously learn from experience during operation. Throughout this prospectus, we show algorithms that can achieve in obtaining these ingredients. First, we will demonstrate a simple algorithm that plans for the most energy-efficient trajectories for a quadruped robot by optimizing for parameters such as cost of transport, manipulability measures, and avoid non-slipping configurations. With this algorithm, we show that the robot only moves when necessary, anddemonstrates behaviors of reacting to outside disturbances to ensure it does not fall while also not wasting unnecessary energy. The idea of understanding is demonstrated through an algorithm that combines an MPC, SLAM, RNN, and object detection using CNNs to generate paths for unknown and uncertain environments. This algorithm is evaluated not only for a complex quadruped robot, but also for multi-agent robot teams consisting of a UGV and UAV. The feasibility of such complex algorithm is also evaluated. Lastly, the idea of continuous learning is addressed not only through use of learning-based algorithms such as RNNs, but also through auto-tuning algorithms that employ an UKF. Using a UKF, we show that we can automatically tune controller gains and even parameters of an online planner. Because the UKF can adapt parameters quickly and without heavy computational load, the robot can continuously adapt its control/planner parameters during online operation to continually learn from the environment. To summarize, we will first present a simple planner for energy-efficient locomotion, provide two examples of end-to-end frameworks for motion planning and state estimation that uses a hybrid approach consisting of model and learning-based methods, and then provide a method of calibrating such end-to-end frameworks (which often contain many various modules) through an auto-tuning technique. Lastly, I end with a discussion on Large Language Models, and how they may potentially affect the robotic field, and further contribute to the idea of understanding in significant ways.