Toward Safety Guarantees For Fully-Actuated Robots

Abstract

Ensuring the safe operation of robots is critical to fostering public trust and enabling widespread deployment. However, safety is challenged by two factors: the computational burden of trajectory generation for high degree-of-freedom (DoF) systems and the difficulty of trajectory tracking under model uncertainties. Variability from manufacturing, assembly, wear, and the environment can cause robots with identical designs to behave differently under identical inputs. This dissertation presents a unified framework to address these challenges, ensuring safe and reliable control of fully actuated robots.

First, a novel, efficient trajectory optimization method is introduced for constrained fully actuated systems. By parameterizing trajectories using Bezier curves on actuated joints and reconstructing the full state via inverse dynamics, this approach reduces the number of decision variables and constraints, resulting in significantly faster trajectory synthesis. Its efficacy is demonstrated on humanoid robots with closed-loop kinematic constraints.

Second, a robust control method is developed to guarantee tracking error bounds under model uncertainty. While conventional controllers emphasize asymptotic convergence, they often lack explicit robustness to parameter variation—particularly relevant when manipulating unknown payloads. The proposed controller treats parameter uncertainty as bounded disturbances and is proven to guarantee performance and safety. Its effectiveness is validated both in simulation and on hardware platforms.

Third, the dissertation presents a system identification algorithm tailored for humanoid robots with closed-loop kinematics. Existing identification methods cannot directly handle such constraints. The proposed approach guarantees physical consistency of estimated parameters and is validated on a hardware humanoid. Integrating these parameters into the robust controller improves tracking compared to manufacturer-supplied values.

Fourth, a provably-safe online system identification pipeline is developed for manipulators with unknown payloads. Conventional strategies often rely on conservative assumptions about payloads, which can compromise performance. This pipeline generates exciting yet safe trajectories—maximizing information about inertial parameters—while respecting joint, torque, and collision constraints. Using the collected data, a novel identification method provides a conservative overapproximation of payload parameters, which is then used by the robust controller to maintain safety and optimize performance.

In summary, this dissertation introduces an integrated framework for safe and efficient control of fully actuated robotic systems under uncertainty. It contributes novel methods for fast trajectory generation, robust tracking with safety guarantees, and physically consistent system identification for complex robotic platforms. The framework is validated through hardware and simulation experiments on manipulators and humanoids, enhancing the safety and reliability of robotic systems in real-world applications.