Robotic introspection and online decision making have been an area ofincreased focus. The goal is to endow robots with the ability to understandtheir actions and make timely decisions to reach their goals. Particularly, inunstructured environments, external perturbations are hard to model inlow-level control systems and often lead to failure. Robots must thenunderstand nominal and anomalous conditions and trigger timely responses tobehaviors that allow the robot to recover and even learn from them and preventthem. Our contribution is the implementation of a fast and robust robotintrospection system that allows recovery from (one or multiple) anomaloussituations at any point in the task. The system handles both internal modelingerrors as well as external perturbations. The robustness of the system isdemonstrated across multiple manipulation tasks. The system assumes tasks aredecomposed into a sequence of nodes, where each node performs a dual role: oneof motion generation and one of introspection. Motion generation is flexibleand can be done with any type of accessible approach. Introspection is done bymodeling the robots multimodal signals using a range of HMMs includingnonparametric Bayesian hidden Markov models. Such models yield strongexpressive power to discriminate both nominal and anomalous situations. We madeuse of a generic strategy for recovery that is easy and flexible to designacross different tasks. A new metric for anomaly detection, critical in theproper assessment of the system after recovery has taken place was alsodesigned. We show how the system recovers from both pose estimation errors thatlead to collisions in pick tasks as well as external human collisions.Furthermore, the system is able to robustly recover from collisions that occurat multiple points in the task; even, when anomalies repeatedly appear at aspecific point in the task.
展开▼
