Reinforcement learning (RL) algorithms should learn as much as possible about the environment but not the properties of the physics engines that generate the environment. There are multiple algorithms that solve the task in a physics engine based environment but there is no work done so far to understand if the RL algorithms can generalize across physics engines. In this work, we compare the generalization performance of various deep reinforcement learning algorithms on a variety of control tasks. Our results show that MuJoCo is the best engine to transfer the learning to other engines. On the other hand, none of the algorithms generalize when trained on PyBullet. We also found out that various algorithms have a promising generalizability if the effect of random seeds can be minimized on their performance.
Estimating epistemic uncertainty of models used in low-latency applications and Out-Of-Distribution samples detection is a challenge due to the computationally demanding nature of uncertainty estimation techniques. Estimating model uncertainty using approximation techniques like Monte Carlo Dropout (MCD), DropConnect (MCDC) requires a large number of forward passes through the network, rendering them inapt for low-latency applications. We propose Select-DC which uses a subset of layers in a neural network to model epistemic uncertainty with MCDC. Through our experiments, we show a significant reduction in the GFLOPS required to model uncertainty, compared to Monte Carlo DropConnect, with marginal trade-off in performance. We perform a suite of experiments on CIFAR 10, CIFAR 100, and SVHN datasets with ResNet and VGG models. We further show how applying DropConnect to various layers in the network with different drop probabilities affects the networks performance and the entropy of the predictive distribution.
Facial emotion recognition is the task to classify human emotions in face images. It is a difficult task due to high aleatoric uncertainty and visual ambiguity. A large part of the literature aims to show progress by increasing accuracy on this task, but this ignores the inherent uncertainty and ambiguity in the task. In this paper we show that Bayesian Neural Networks, as approximated using MC-Dropout, MC-DropConnect, or an Ensemble, are able to model the aleatoric uncertainty in facial emotion recognition, and produce output probabilities that are closer to what a human expects. We also show that calibration metrics show strange behaviors for this task, due to the multiple classes that can be considered correct, which motivates future work. We believe our work will motivate other researchers to move away from Classical and into Bayesian Neural Networks.
Deep learning models are extensively used in various safety critical applications. Hence these models along with being accurate need to be highly reliable. One way of achieving this is by quantifying uncertainty. Bayesian methods for UQ have been extensively studied for Deep Learning models applied on images but have been less explored for 3D modalities such as point clouds often used for Robots and Autonomous Systems. In this work, we evaluate three uncertainty quantification methods namely Deep Ensembles, MC-Dropout and MC-DropConnect on the DarkNet21Seg 3D semantic segmentation model and comprehensively analyze the impact of various parameters such as number of models in ensembles or forward passes, and drop probability values, on task performance and uncertainty estimate quality. We find that Deep Ensembles outperforms other methods in both performance and uncertainty metrics. Deep ensembles outperform other methods by a margin of 2.4% in terms of mIOU, 1.3% in terms of accuracy, while providing reliable uncertainty for decision making.
Fast estimates of model uncertainty are required for many robust robotics applications. Deep Ensembles provides state of the art uncertainty without requiring Bayesian methods, but still it is computationally expensive. In this paper we propose deep sub-ensembles, an approximation to deep ensembles where the core idea is to ensemble only the layers close to the output, and not the whole model. With ResNet-20 on the CIFAR10 dataset, we obtain 1.5-2.5 speedup over a Deep Ensemble, with a small increase in error and NLL, and similarly up to 5-15 speedup with a VGG-like network on the SVHN dataset. Our results show that this idea enables a trade-off between error and uncertainty quality versus computational performance.
Marine and Underwater resources are important part of the economy of many countries. This requires significant financial resources into their construction and maintentance. Robotics is expected to fill this void, by automating and/or removing humans from hostile environments in order to easily perform maintenance tasks. The Robocademy Marie Sklodowska-Curie Initial Training Network was funded by the European Union's FP7 research program in order to train 13 Fellows into world-leading researchers in Marine and Underwater Robotics. The fellows developed guided research into three areas of key importance: Autonomy, Disturbance Rejection, and Perception. This paper presents a summary of the fellows' research in the three action lines. 71 scientific publications were the primary result of this project, with many other publications currently in the pipeline. Most of the fellows have found employment in Europe, which shows the high demand for this kind of experts. We believe the results from this project are already having an impact in the marine robotics industry, as key technologies are being adopted already.
Traffic sign recognition is an important component of many advanced driving assistance systems, and it is required for full autonomous driving. Computational performance is usually the bottleneck in using large scale neural networks for this purpose. SqueezeNet is a good candidate for efficient image classification of traffic signs, but in our experiments it does not reach high accuracy, and we believe this is due to lack of data, requiring data augmentation. Generative adversarial networks can learn the high dimensional distribution of empirical data, allowing the generation of new data points. In this paper we apply pix2pix GANs architecture to generate new traffic sign images and evaluate the use of these images in data augmentation. We were motivated to use pix2pix to translate symbolic sign images to real ones due to the mode collapse in Conditional GANs. Through our experiments we found that data augmentation using GAN can increase classification accuracy for circular traffic signs from 92.1% to 94.0%, and for triangular traffic signs from 93.8% to 95.3%, producing an overall improvement of 2%. However some traditional augmentation techniques can outperform GAN data augmentation, for example contrast variation in circular traffic signs (95.5%) and displacement on triangular traffic signs (96.7 %). Our negative results shows that while GANs can be naively used for data augmentation, they are not always the best choice, depending on the problem and variability in the data.
Detecting novel objects without class information is not trivial, as it is difficult to generalize from a small training set. This is an interesting problem for underwater robotics, as modeling marine objects is inherently more difficult in sonar images, and training data might not be available apriori. Detection proposals algorithms can be used for this purpose but usually requires a large amount of output bounding boxes. In this paper we propose the use of a fully convolutional neural network that regresses an objectness value directly from a Forward-Looking sonar image. By ranking objectness, we can produce high recall (96 %) with only 100 proposals per image. In comparison, EdgeBoxes requires 5000 proposals to achieve a slightly better recall of 97 %, while Selective Search requires 2000 proposals to achieve 95 % recall. We also show that our method outperforms a template matching baseline by a considerable margin, and is able to generalize to completely new objects. We expect that this kind of technique can be used in the field to find lost objects under the sea.
Garbage and waste disposal is one of the biggest challenges currently faced by mankind. Proper waste disposal and recycling is a must in any sustainable community, and in many coastal areas there is significant water pollution in the form of floating or submerged garbage. This is called marine debris. Submerged marine debris threatens marine life, and for shallow coastal areas, it can also threaten fishing vessels [I\~niguez et al. 2016, Renewable and Sustainable Energy Reviews]. Submerged marine debris typically stays in the environment for a long time (20+ years), and consists of materials that can be recycled, such as metals, plastics, glass, etc. Many of these items should not be disposed in water bodies as this has a negative effect in the environment and human health. This thesis performs a comprehensive evaluation on the use of DNNs for the problem of marine debris detection in FLS images, as well as related problems such as image classification, matching, and detection proposals. We do this in a dataset of 2069 FLS images that we captured with an ARIS Explorer 3000 sensor on marine debris objects lying in the floor of a small water tank. The objects we used to produce this dataset contain typical household marine debris and distractor marine objects (tires, hooks, valves, etc), divided in 10 classes plus a background class. Our results show that for the evaluated tasks, DNNs are a superior technique than the corresponding state of the art. There are large gains particularly for the matching and detection proposal tasks. We also study the effect of sample complexity and object size in many tasks, which is valuable information for practitioners. We expect that our results will advance the objective of using Autonomous Underwater Vehicles to automatically survey, detect and collect marine debris from underwater environments.