APPLD: Adaptive Planner Parameter Learning From Demonstration

Abstract

While current autonomous navigation systems allow robots to successfully drive themselves from one point to another in specific environments, they typically require extensive manual parameter re-tuning by human robotics experts in order to function in new environments. Furthermore, even for just one complex environment, a single set of fine-tuned parameters may not work well in different regions of that environment. These problems prohibit reliable mobile robot deployment by nonexpert users. As a remedy, we propose Adaptive Planner Parameter Learning (APPL), a machine learning framework that can leverage non-expert human interaction via several modalitiesincluding teleoperated demonstrations, corrective interventions, and evaluative feedback-and also unsupervised reinforcement learning to learn a parameter policy that can dynamically adjust the parameters of classical navigation systems in response to changes in the environment. APPL inherits safety and explainability from classical navigation systems while also enjoying the benefits of machine learning, i.e., the ability to adapt and improve from experience. We present a suite of individual APPL methods and also a unifying cycle-of-learning scheme that combines all the proposed methods in a framework that can improve navigation performance through continual, iterative human interaction and simulation training.

AutoTuning Environment for Static Obstacle Avoidance Methods Applied to USVs

“…where d wp and d collision represent the distance thresholds, which define whether the USV has reached P goal or has collided, respectively. In this way, if the parameter vector Θ tuning drives the USV to the goal, within the established limits (39), the index J s is used to minimize: time, distance and control effort in the travel, see Equation (38). On the other hand, if the safety distance (d collision ) or any of the established limits is exceeded, the numerical simulation is stopped and the index used is J f .…”

“…This last facilitates that, if in the initial populations a parameter vector that successfully guides the USV to P goal is not obtained, the GA prioritizes those vectors Θ tuning that drive the USV closest to the goal. Once the fitness function (38) has been described, the tuning parameter vectors of the four SOA methods studied in this work are defined in the Equation (40).…”

“…In this way, the main contribution of this work is achieved, a new autotuning environment for SOA methods applied to USVs. Currently, there are different approaches for the autotuning of unmanned vehicles [38][39][40][41][42][43][44]. In [38] the authors use machine learning for autotuning the guidance system on a physical ground vehicle.…”

“…Currently, there are different approaches for the autotuning of unmanned vehicles [38][39][40][41][42][43][44]. In [38] the authors use machine learning for autotuning the guidance system on a physical ground vehicle. As autotuning methods applied in the marine environment, in [39] a genetic algorithm is used to autotune the parameters of a sliding mode controller, which is applied to the mathematical model of an oil tanker vessel.…”

AutoTune: Controller Tuning for High-Speed Flight

“…If demonstrations by a human expert are available, another option consists of using inverse reinforcement learning [17], [18], but this is generally not the case with time-optimal trajectories, which can be faster than the trajectories flown by the best human pilots [1].…”