Do Not Make the Same Mistakes Again and Again: Learning Local Recovery Policies for Navigation From Human Demonstrations

Abstract

In this paper, we present a human-in-the-loop learning framework for mobile robots to generate effective local policies in order to recover from navigation failures in longterm autonomy. We present an analysis of failure and recovery cases derived from long-term autonomous operation of a mobile robot, and propose a two-layer learning framework that allows to detect and recover from such navigation failures. Employing a learning by demonstration (LbD) approach, our framework can incrementally learn to autonomously recover from situations it initially needs humans to help with. The learning framework allows for both real-time failure detection and regression using Gaussian processes (GPs). Our empirical results on two different failure scenarios indicate that given 40 failure state observations, the true positive rate of the failure detection model exceeds 90%, ending with successful recovery actions in more than 90% of all detected cases.

Taxonomy of Trust-Relevant Failures and Mitigation Strategies

“…Rich data sets, comprising task and error logs [32], user demographics [36], and navigation failures [15] have been obtained from these deployments, and analysed for the case study for this paper.…”

“…Probability and Detection. In the specific instance, a variety of problems were detected automatically, such as navigation issues [15], forceful pushes to the robot, and hardware failures. Consequently, many failure types can be detected from system logs and from dedicated anomaly or failure detection modules that allow to estimate the probability of them occurring.…”

“…An analysis of software failures, in particular navigation failures (which account for more than 99% of all software-related issues in this particular use case) in [15] reveals that in 1605 instances the robot had to ask for help as it could not recover from a navigation problem, making its failure obvious to the interacting humans, and hence potentially having an impact on trust. Thus, we observed such failure about every 2 hours of autonomous operation, leading to a ratio of 7 : 1 for Software failures to interactions, leading to a Probability score of "3" in Tab.…”

Introduction to the Special Issue on AI for Long-Term Autonomy

“…An example of the latter is the work by Del Duchetto et al [2], who employ Gaussian Processes in order to interactively learn local navigation recovery behaviours for mobile robots. Using a long-term dataset, they highlight that navigation failures are predictably localised in space-time, and propose a solution that can exploit rare interactive teachin opportunities from situations where the robot was helped by human bystanders.…”

Engaging Learners in Dialogue Interactivity Development for Mobile Robots

“…The work we present in this paper is part of a wider research programme concerned with improving the long-term autonomy of mobile robots in humancentred environments, as an extension of prior and continuing research [7,4]. Of specific interest to the current effort is to increase the social interactivity of the robots concerned so as to reduce barriers to interaction and increase utility for non-technical users.…”