Safe Visual Navigation via Deep Learning and Novelty Detection

By Charles Richter, Nicholas A. Roy. 2017

202 citations recorded by the source corpus.

Robots that use learned perceptual models in the real world must be able to safely handle cases where they are forced to make decisions in scenarios that are unlike any of their training examples. However, state-of-the-art deep learning methods are known to produce erratic or unsafe predictions when faced with novel inputs. Furthermore, recent ensemble, bootstrap and dropout methods for quantifying neural network uncertainty may not efficiently provide accurate uncertainty estimates when queried with inputs that are very different from their training data. Rather than unconditionally trusting the predictions of a neural network for unpredictable real-world data, we use an autoencoder to recognize when a query is novel, and revert to a safe prior behavior. With this capability, we can deploy an autonomous deep learning system in arbitrary environments, without concern for whether it has received the appropriate training. We demonstrate our method with a vision-guided robot that can leverage its deep neural network to navigate 50% faster than a safe baseline policy in familiar types of environments, while reverting to the prior behavior in novel environments so that it can safely collect additional training data and continually improve. A video illustrating our approach is available at: http://groups.csail.mit.edu/rrg/videos/safe visual navigation.

Subfields: Artificial Intelligence, Robotics & Control

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Robotics & Control

Citations in the corpus

Deep Drone Racing: Learning Agile Flight in Dynamic Environments (2018, background): “While being independent of any global map and position estimate, these methods are not directly applicable to our specific problem due to their high computational complexity [27, 11], their low maximum speed [13] or the inherent difficulties of generalizing to 3D motions [25, 4, 24].”
Self-Supervised Deep Reinforcement Learning with Generalized Computation Graphs for Robot Navigation (2017, methodology): “These supervised learning methods include learning: drivable routes and then using a planner [13], near-to-far obstacle detectors [14], reactive controllers on top of a map-based planner [15], driving affor-”
Composable Action-Conditioned Predictors: Flexible Off-Policy Learning for Robot Navigation (2018, background): “Much of the reinforcement learning work for robot navigation has focused on collision avoidance [27, 28, 15], while our work addresses goal-directed, multi-objective navigation.”
Leave no Trace: Learning to Reset for Safe and Autonomous Reinforcement Learning (2017, positive): “However, in contrast to our approach, Richter & Roy (2017) use a heuristic, hand-engineered reset policy, while our reset policy is learned simultaneously with the forward policy.”
Mo' States Mo' Problems: Emergency Stop Mechanisms from Observation (2019, background): “Richter and Roy [19] use an auto-encoder as an anomaly detector to determine when a high dimensional state is anomalous, and revert to a safe policy.”
Variational Autoencoder for End-to-End Control of Autonomous Driving with Novelty Detection and Training De-biasing (2018, background): “Previous works have leveraged VAEs for novelty detection [25], but did not directly use the predicted distributions of the sensor for a model fit criterion.”
DroNet: Learning to Fly by Driving (2018, background): “Moreover, to make our system more robust, one could produce a measure of uncertainty, as in [28].”
A Survey on Active Simultaneous Localization and Mapping: State of the Art and New Frontiers (2022, background): “How does scene prediction behave in unstructured environments? How to account for uncertainty? Is measurement prediction more reliable and informative than map prediction? How to predict the effect of only a certain set of nonmyopic actions in the map rather than augmenting the whole scene? Regardin”
GONet: A Semi-Supervised Deep Learning Approach For Traversability Estimation (2018, methodology): “Our experiments show that the proposed methods outperform supervised baselines, and that small amounts of negative examples can improve traversability estimation in comparison to using only positive data, as in [23].”
A survey of uncertainty in deep neural networks (2021, background): “Notably, [323] proposed to perform novelty detection using auto-encoders, where the reconstructed outputs of auto-encoders was used to decide how much one can trust the network’s predictions.”
Deep One-Class Classification (2018, background): “…& Frey, 2013), variational autoencoders (VAEs) (Kingma & Welling, 2013), and deep convolutional autoencoders (DCAEs) (Masci et al., 2011; Makhzani & Frey, 2015), where the last variant is predominantly used in AD applications with image or video data (Seeb¨ock et al., 2016; Richter & Roy, 2017).”
Learning Novel Policies For Tasks (2019, methodology): “To generalize across a variety of problems, using neural network for novelty detection as in (Hawkins et al., 2002; Richter & Roy, 2017; Ruff et al., 2018) becomes a more desired approach.”
Following High-level Navigation Instructions on a Simulated Quadcopter with Imitation Learning (2018, background): “Our goal is orthogonal to providing safety guarantees [15, 22, 43].”
RoboTHOR: An Open Simulation-to-Real Embodied AI Platform (2020, background): “Spatial navigation approaches [62, 17, 64, 23, 25, 56, 79, 49, 30, 14, 10] typically address navigating towards a pre-specified coordinate or a frame of a scene and they focus on understanding the geometry of the scene and learning better exploration strategies.”
Planning and Decision-Making for Autonomous Vehicles (2018, background): “Therefore, an autoencoder may be trained to recognize when to detect novelties in the input data (80) and revert from an end-to-end approach to a safe non-learning-based behavior, such as the execution of conventional motion primitives.”
Recovery RL: Safe Reinforcement Learning With Learned Recovery Zones (2020, background): “Richter et al. [27] learns the probability of constraint violation conditioned on an action plan to activate a hand-designed safety controller.”