Forward Prediction of Target Localization Failure Through Pose Estimation Artifact Modelling

Abstract

For safety critical applications the ability of localization systems to self-assess their own performance and know when they are failing is as important as absolute accuracy. Previous methods have self-assessed current system performance, identifying failure after it occurs. We propose to instead pre-emptively avoid failure by predicting likely localization performance at locations not yet explored by a robot. To achieve this, we propose an approach for supervising a target object localization system by modelling trends in internal pipeline artifacts that are predictive of localization accuracy. We use this model to predict where acceptable localization performance is possible and where failure is likely. We evaluate our approach with both off-line recorded datasets and live robot experiments in the context of an upper limb surgical task using human bone phantoms as localization targets. We demonstrate our approach implemented as both a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Long Range Predictor</i> for use in informing future planning, and a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Next-Step Predictor</i> , for ongoing task supervision to stop a robot before reaching localization failure. We show that our method provides significant improvement over a naive baseline achieving a mean increase in safe path length, or usable workspace, without localization failure of 84.1% for our long range predictor and 102.1% for our next-step predictor.