Quantifying and Maximizing Performance of a Human-Centric Robot under Precision, Safety, and Robot Specification Constraints

The goal of this research is to formulate new theories and techniques to achieve 99.00% safe mobile robot manipulators in human-robot-interaction (HRI) scenarios without compromising performance by integrating human-awareness in the low-level control loop and high-level decision process of the robot using multi-modal sensing for intent recognition, planning for action generation, and compliant control for unintended collisions. The research thrusts can be categorized into four areas: (1) human detection, (2) human intent recognition, (3) planning & action generation, and (4) whole-body compliant control. Throughout the fellowship, significant contributions have been made in research thrusts (3) and (4) in the context of humanoid robots, while only minor contributions were made on thrusts (1) and (2). For thrust (1), a human detection algorithm using a LIDAR sensor was developed and was competitive to other LIDAR-based approaches to detecting humans prior to the widespread use of convolutional-neural networks (CNN). When CNN approaches with RGB cameras became the de-facto standard for detecting humans, a Robot-Operating System (ROS) wrapper was developed and released as an open-sourced software using the best-performing CNN-based approach. The software has garnered a modest following and has been used as a gateway to using CNN-based human detection algorithms with robotic systems. For thrust (2), a gesture recognition algorithm was created and a novel approach to modeling humans as dynamical system was explored. For the former, gestures were parameterized as dynamic-movement primitives and classified using gaussian mixture models. Again, prior to using neural networks, these methods were competitive. For the latter, human behavior and their interaction with robotic systems were modeled as a linear dynamical system, dubbed as HRI-dynamical systems. With a mathematical model of HRI systems, a model-predictive control (MPC) approach was used to generate control policies for the robotic system. HRI systems as linear dynamical systems are a novel approach to modeling human behavior and their interaction with robots. Provided that the interaction can be modeled as a dynamical system, our work shows that optimal control techniques can be utilized to generate optimal robot policies. For thrust (3), novel action generation and planning approaches were discovered by combining optimization and search techniques in addition to the aforementioned MPC approach. First, a novel thermal recovery algorithm was developed by giving actuator thermal awareness to a limbed-robotic system. With a minimization approach, this enabled the robot to take actions to recover thermally compromised actuators while satisfying balancing constraints. This pushes the state-of-art by integrating self-thermal-awareness as part of the robot’s planning process. This was also the first demonstration of the concept of thermal inverse kinematics (Thermal IK). Next, to increase the efficiency of robotic manipulators, particularly with limbs, a novel locomanipulation planner was developed. By defining a locomotion constraint manifold a desired end-effector path, feasible manipulation trajectories that are admissible in the constraint manifold can be searched. This is a novel and state-of-art approach in generating simultaneous locomotion and manipulation trajectories for limbed systems. For thrust (4), novel control methodologies were developed to improve whole-body compliant control of robotic systems. First, a control design approach towards tuning series-elastic actuators (SEAs) was developed to improve either the impedance or force control behavior of the SEA. A sensitivity analysis was performed to show the performance of the controller and the concept of Z-region is introduced to quantify the improvement. This increases the low-level performance of torque-controlled robots enabling better whole-body compliant control and enhancing the safety of physical human-robot interaction. Second, a novel configuration-based whole-body controller was developed to enable passive-ankled biped systems to better locomote in unstructured terrain. This new whole-body controller enables the system to be robust to dynamic model inaccuracies which degrade the whole-body dynamic-based controllers. Using kinematic information to generate joint feedback commands and the dynamic model only for feed-forward torques, the unique structure exploitation enables improved whole-body control and was verified with a highly underactuated robotic system. Simulations also verify that the new whole-body control approach works with other robotic systems as well, proving that the information structure exploitation is retained for general robotic systems.