Learning Deep Generative Spatial Models For Mobile Robots

II. RELATED WORKRepresenting semantic spatial knowledge is a broadlyresearched topic, with many solutions employing vision [17][18][2][5]. Images clearly carry rich informationabout semantics; however, they are also affected by changing environment conditions. At the same time, roboticsresearchers have seen advantages of using range data thatare much more robust in real-world settings and easier toprocess in real time. In this work, we focus on laser-rangedata, as a way of introducing and evaluating a new spatialmodel as well as a recently proposed deep architecture.Laser-range data have been extensively used for placeclassification and semantic mapping, and many traditional,handcrafted representations have been proposed. Buschkaet al. [19] contributed a simple method that incrementallydivided grid maps of indoor environments into two classesof open spaces (rooms and corridors). Mozos et al. [12]applied AdaBoost to create a classifier based on a set ofmanually designed geometrical features to classify placesinto rooms, corridors and doorways. In [20], omnidirectionalvision was combined with laser data to build descriptors,called fingerprints of places. Finally, in [13], SVMs havebeen applied to the geometrical features of Mozos et al. [12]leading to significant improvement in performance over theoriginal AdaBoost. That approach has been further integratedwith visual and object cues for semantic mapping in [2].Deep learning and unsupervised feature learning, aftermany successes in speech recognition and computer vision [3], entered the field of robotics with superior performance in object recognition [21][22] and robot grasping [23][4]. The latest work in place classification alsoemploys deep approaches. In [5], deep convolutional network(CNN) complemented with a series of one-vs-all classifiers isused for visual semantic mapping. In [6], CNNs are used toclassify grid maps built from laser data into 3 classes: room,corridor, and doorway. In these works, deep models are usedexclusively for classification, and use of generative modelshas not been explored. In contrast, we propose a universalprobabilistic generative model, and demonstrate its usefulness for multiple robotics tasks, including classification.Several generative, deep architectures have recently beenproposed, notably Variational Autoencoders [24], Generative Adversarial Networks [14], and Sum-Product Networks[8][7][9]. GANs have been shown to produce high-qualitygenerative representations of visual data [15], and have beensuccessfully applied to image completion [25]. SPNs, aprobabilistic model, achieved promising results for variedapplications such as speech [10] and language modeling [26],human activity recognition [11], and image classification [9]and completion [7], but have not been used in robotics. In thiswork, we exploit the universality and efficiency of SPNs topropose a single spatial model able to solve a wide range ofinference problems relevant to a mobile robot. Furthermore,inspired by their results in other domains, we also evaluateGANs (when applicable). This serves as a comparison anda demonstration of GANs on a new application.Fig. 1: An SPN for a naive Bayes mixture model P (X1 , X2 ),with three components over two binary variables. The bottomlayer consists of indicators for X1 and X2 . Weighted sumnodes, with weights attached to inputs, are marked with ,while product nodes are marked with . Y1 represents alatent variable marginalized out by the root sum.III. SUM-PRODUCT NETWORKSSum-product networks are a recently proposed probabilistic deep architecture with several appealing properties andsolid theoretical foundations [8][7][9]. One of the primarylimitations of probabilistic graphical models is the complexity of their partition function, often requiring complex approximate inference in the presence of non-convex likelihoodfunctions. In contrast, SPNs represent joint or conditionalprobability distributions with partition functions that areguaranteed to be tractable and involve a polynomial numberof sum and product operations, permitting exact inference.SPNs are a deep, hierarchical representation, capable ofrepresenting context-specific independence and performingfast, tractable inference on high-treewidth models. While notall probability distributions can be encoded by polynomialsized SPNs, recent experiments in several domains show thatthe class of distributions modeled by SPNs is sufficient formany real-world problems, offering real-time efficiency.As shown in Fig. 1, on a simple example of a naive Bayesmixture model, an SPN is a generalized directed acyclicgraph composed of weighted sum and product nodes. Thesums can be seen as mixture models over subsets of variables, with weights representing mixture priors. Products canbe viewed as features or mixture components. The latentvariables of the mixtures can be made explicit and theirvalues inferred. This is often done for classification models,where the root sum is a mixture of sub-SPNs representingclasses. The bottom layers effectively define features reactingto certain values of indicators1 for the input variables.Formally, following Poon & Domingos [7], we can definean SPN as follows:Definition 1: An SPN over variables X1 , . . . , XV is arooted directed acyclic graph whose leaves are the indicators(X11 , . . . , X1I ), . . . , (XV1 , . . . , XVI ) and whose internal nodesare sums and products. Each edge (i, j) emanating from asum node i has a non-negative weight wij . The value of aproduct node is the product of the values of its children. The1 Indicator is a binary variable set to 1 when the corresponding categoricalvariable takes a specific value. For using continuous input variables, see [7].756

Pvalue of a sum node is j Ch(i) wij vj , where Ch(i) are thechildren of i and vj is the value of node j. The value of anSPN S[X1 , . . . , XV ] is the value of its root.Not all architectures consisting of sums and products resultin a valid probability distribution. However, following simpleconstraints on the structure of an SPN will guarantee validity(see [7], [8]). When the weights of each sum node are normalized to sum to 1, the value of a valid SPN S[X11 , . . . , XVI ]is equal to the normalized probability P (X1 , . . . , XV ) of thedistribution modeled by the network [8].A. Generating SPN structureThe structure of the SPN determines the group of distributions that can be learned. Therefore, most previousworks [9][11][26] relied on domain knowledge to design theappropriate structure. Furthermore, several structure learningalgorithms were proposed [27][28] to discover independencies between the random variables in the dataset, andstructure the SPN accordingly. In this work, we experimentwith a different approach, originally hinted at in [7], whichgenerates a random structure, as in random forests. Such anapproach has not been previously evaluated. Our experimentsdemonstrate that it can lead to very good performance andcan accommodate a wide range of distributions. Additionally,after parameter learning, the generated structure can bepruned by removing edges associated with weights close tozero. This can be seen as a form of structure learning.To obtain the random structure, we recursively generatenodes based on multiple random decompositions of a setof random variables into multiple subsets until each subsetis a singleton. As illustrated in Fig. 2 (middle, Level 1),at each level the current set of variables to be decomposedis modeled by multiple mixtures (green nodes), and eachsubset of the decomposition is also modeled by multiplemixtures (green nodes one level below). Product nodes(blue) are used as an intermediate layer and act as featuresdetecting particular combinations of mixtures representingeach subset. The top mixtures of each level mix outputs ofall product nodes at that level. The same set of variables canbe decomposed into subsets in multiple random ways (e.g.there are two decompositions at the top of Fig. 2).B. Inference and LearningInference in SPNs is accomplished by a single passthrough the network. Once the indicators are set to representthe evidence, the upward pass will yield the probability of theevidence as the value of the root node. Partial evidence (ormissing data) can easily be expressed by setting all indicatorsfor a variable to 1. Moreover, since SPNs compute a networkpolynomial [29], derivatives computed over the network canbe used to perform inference for modified evidence withoutrecomputing the whole SPN. Finally, it can be shown [8] thatMPE inference in a certain class of SPNs (selective) can beperformed by replacing all sum nodes with max nodes whileretaining the weights. Then, the indicators of the variablesfor which the MPE state is inferred are all set to 1 and astandard upward pass is performed. A downward pass thenFig. 2: The structure of the SPN implementing our spatialmodel. The bottom images illustrate a robot in an environment and a robot-centric polar grid formed around the robot.The SPN is built on top of the variables representing theoccupancy in the polar grid.follows, which recursively selects the highest valued childof each sum (max) node and all children of a product node.The indicators selected by this process indicate the MPE stateof the variables. In general SPNs, this algorithm yields anapproximation of the MPE state.SPNs lend themselves to be learned generatively [7] ordiscriminatively [9] using Expectation Maximization (EM) or757

(a) Corridor(b) Doorway(c) Small Office(d) Large Officewill be built by integrating multiple models of local places.Thus, we constrain the observation of a place to the information visible from the robot (structures that can be raytracedfrom the robot’s location). As a result, walls occlude the viewand the local map mostly contains information from a singleroom. In practice, additional noise is almost always present,but is averaged out during learning of the model. Examplesof such local environment observations can be seen in Fig. 3.Next, each local observation is transformed into a robotcentric polar occupancy grid (compare polar and Cartesiangrids in Fig. 3). The resulting observation contains higherresolution details closer to the robot and lower-resolutioncontext further away. This focuses the attention of the modelon the nearby objects. Higher resolution of informationcloser to the robot is important for understanding the semantics of the robot’s exact location (for instance when therobot is at a doorway). However, it also relates to how spatialinformation is used by a mobile robot when planning andexecuting actions. It is in the vicinity of the robot that higheraccuracy of spatial information is required. A similar principle is exploited by many navigation components, which usedifferent resolution of information for local and global pathplanning. Additionally, such a representation corresponds tothe way the robot perceives the world because of the limitedresolution of its sensors. Our goal is to use a similar strategywhen representing 3D and visual information, by extendingthe polar representation to 3 dimensions. Finally, a highresolution map of the complete environment can be largelyrecovered by integrating stored or inferred polar observationsover the path of the robot. We built polar grids of radius of5m, with an angle step of 6.4 degrees and grid resolutiondecreasing with the distance from the robot.Fig. 3: Local environment observations used in our experiments, expressed as Cartesian and polar occupancy grids, forexamples of places of different semantic categories.gradient descent. In this work, we employ hard EM to learnthe weights, which was shown to work well for generativelearning [7]. As is often the case for deep models, the gradient quickly diminishes as more layers are added. Hard EMovercomes this problem, permitting learning of SPNs withhundreds of layers. Each iteration of the EM learning consistsof an MPE inference of the implicit latent variables of eachsum with training samples set as evidence (E step), and anupdate of weights based on the inference results (M step,for details, see [7]). We achieved best results by modifyingthe MPE inference to use sums instead of maxes during theupwards pass, while selecting the max valued child duringthe downward pass. Furthermore, we performed additivesmoothing when updating the weights corresponding to aDirichlet prior and terminated learning after 300 iterations.No additional learning parameters are required.B. Architecture of DGSMThe architecture of DGSM is based on a generative SPNillustrated in Fig. 2. The model learns a probability distribution P (Y, X1 , . . . , XC ), where Y represents the semanticcategory of a place, and X1 , . . . , XC are input variables representing the occupancy in each cell of the polar grid. Eachoccupancy cell is represented by three indicators in the SPN(for empty, occupied and unknown space). These indicatorsconstitute the bottom of the network (orange nodes).The structure of the model is partially designed basedon domain knowledge and partially generated according tothe algorithm described in Sec. III-A. The resulting modelis a single SPN assembled from three levels of sub-SPNs.We begin by splitting the polar grid equally into 8 45degree views. For each view, we generate a random subSPN by recursively building a hierarchy of decompositionsof subsets of polar cells in the view. Then, on top of allthe sub-SPNs representing the views, we generate an SPNrepresenting complete place geometries for each place class.Finally, the sub-SPNs for place classes are combined by asum node forming the root of the network. The latent variableassociated with that sum node is made explicit as Y andrepresents the semantic class of a place.IV. DEEP GENERATIVE SPATIAL MODEL (DGSM)A. Representing Local EnvironmentsDGSM is designed to support real-time, online spatialreasoning on a mobile robot. A real robot almost always hasaccess to a stream of observations of the environment. Thus,as the first step, we perform spatio-temporal integration ofthe sensory input. We rely on laser-range data, and use aparticle-filter grid mapping [30] to maintain a robot-centricmap of 5m radius around the robot. During acquisition of thedataset used in our experiments, the robot was navigating afixed path through a new environment, while continuouslyintegrating data gathered using a single laser scanner with180 FOV. This still results in partial observations of the surroundings (especially when the robot enters a new room), buthelps to assemble a more complete representation over time.Our goal is to model the geometry and semantics of a localenvironment only. We assume that larger-scale spatial model758

Sub-dividing the representation into views allows us touse networks of different complexity for representing lowerlevel view features and high-level structure of a place. Inour experiments, when representing views, we recursivelydecomposed the set of polar cells using a single randomdecomposition, into 2 cell sub-sets, and generated 4 mixturesto model each such subset. This procedure was repeateduntil each subset contained a single variable representing asingle cell. To increase the discriminative power of each viewrepresentation, we used 14 sums at the top level of the viewsub-SPN. These sums are considered input to a randomlygenerated SPN structure representing a place class. To ensurethat each class can be associated with a rich assortmentof place geometries, we increased the complexity of thegenerated structure and performed 4 random decompositionsof the sets of mixtures representing views into 5 subsets.The performance of the model does not vary greatly withthe structure parameters as long as the generated structure issufficiently expressive to support learning of dependenciesin the data.Several straightforward modifications to the architecturecan be considered. First, the latent variables in the mixturesmodeling each view can be made explicit and considered aview or scene descriptor discovered by the learning algorithm. Second, the weights of the network could be sharedacross views, potentially simplifying the learning process.idea behind GANs is to simultaneously train two deepnetworks: a generative model G(z; θg ) that captures thedata distribution and a discriminative model D(x; θd ) thatdiscriminates between samples from the training data andsamples generated by G. The training alternates betweenupdating θd to correctly discriminate between the true andgenerated data samples and updating θg so that D is fooled.The generator is defined as a function of noise variablesz, typically distributed uniformly (values from -1 to 1 inour experiments). For every value of z, a trained G shouldproduce a sample from the data distribution.Although GANs have been known to be unstable totrain, several architectures have been proposed that resultin stable models over a wide range of datasets. Here, weemploy one such architecture called DC-GAN [15], whichprovides excellent results on datasets such as MNIST, LSUN,ImageNet [15] or CelebA [25]. Specifically, we used 3convolutional layers (of dimensions 18 18 64, 9 9 128,5 5 256) with stride 2 and one fully-connected layer forD2 . We used an analogous architecture based on fractionalstrided convolutions for G. We assumed z to be of size 100.DC-GANs do not use pooling layers, perform batch normalization for both D an G, and use ReLU and LeakyReLUactivations for D and G, respectively. We used ADAM tolearn the parameters.Since DC-GAN is a convolutional model, we could notdirectly use the polar representation as input. Instead, weused the Cartesian local grid maps directly. The resolutionof the Cartesian maps was set to 36 36, which is larger thenthe average resolution of the polar grid, resulting in 1296occupancy values being fed to the DC-GAN, as compared to1176 for DGSM. We encoded input occupancy values into asingle channel2 (0, 0.5, 1 for unknown, empty, and occupied).C. Types of InferenceAs a generative model of a joint distribution betweenlow-level observations and high-level semantic phenomena,DGSM is capable of various types of inferences.First, the model can simply be used to classify observations into semantic categories, which corresponds to MPEinference of y: y argmaxy P (y x1 , . . . , xC ). Second, thelikelihood of an observationP can be used as a measure ofnovelty and thresholded: y P (y, x1 , . . . , xC ) t. We usethis approach to separate test observations of classes knownduring training from observations of unknown classes.If instead, we condition on the semantic information, wecan perform MPE inference over the variables representingoccupancy of polar grid cells:A. Predicting Missing ObservationsIn [25], a method was proposed for applying GANs tothe problem of image completion. The approach first trainsa GAN on the training set and then relies on stochasticgradient descent to adjust the value of z according to a lossfunction L(z) Lc (z) λLp (z), where Lc is a contextualloss measuring the similarity between the generated and trueknown input values, while Lp is a perceptual loss whichensures that the recovered missing values look real to thediscriminator. We use this approach to infer missing observations in our experiments. While effective, it requires iterativeoptimization to infer the missing values. This is in contrastto DGSM, which performs inference using a single up/downpass through the network. We selected the parameter λ toobtain the highest ratio of correctly reconstructed pixels.x 1 , . . . , x C argmax P (x1 , . . . , xC y).x1 ,.,xCThis leads to generation of prototypical examples for eachclass. Finally, we can use partial evidence about the occupancy and infer the most likely state of a subset of polar gridcells for which evidence is missing:XP (y, x1 , . . . , xJ 1 , xJ , . . . , xC )x J , . . . , x C argmaxxJ ,.,xCyVI. EXPERIMENTSWe use this technique to infer missing observations in ourexperiments.We conducted four experiments corresponding to the inference types described in Sec. IV-C. Importantly, DGSM wasV. GANS FOR SPATIAL MODELING2 We evaluated architectures consisting of 4 conv. layers and layers ofdifferent dimensions (depth of the 1st layer ranging from 32 to 256). We alsoinvestigated using two and three channels to encode occupancy information.The final architecture results in significantly better completion accuracy.Recently, Generative Adversarial Networks [14] have received significant attention for their ability to learn complexvisual phenomena in an unsupervised way [15], [25]. The759

(a) DGSM(b) SVM Geometric FeaturesFig. 4: Normalized confusion matrices for the task ofsemantic place categorization.trained only once and the same instance of the model wasused for all inferences. For each experiment, we comparedto a baseline model fine-tuned to the specific sub-problem.Fig. 5: ROC curves for novelty detection. Inliers areconsidered positive, while novel samples are negative.A. Experimental Setupkernel and learning parameters directly on the test sets. Whileearly attempts to solve a similar classification problem withdeep conv nets exist [6], it is not clear whether they offerperformance improvements compared to the SVM-basedapproach. Additionally, using SVMs allows us to evaluatenot only classification, but also novelty detection.The models were trained on the four room categories andevaluated on observations collected in places belonging tothe same category, but on different floors. The normalizedconfusion matrices are shown in Fig. 4. We can see thatDGSM obtains superior results for all classes. The classification rate averaged over all classes (giving equal importanceto each class) and data splits is 85.9% 5.4 for SVM and92.7% 6.2 for DGSM, with DGSM outperforming SVM forevery split. Most of the confusion exists between the smalland large office classes. Offices in the dataset often havecomplex geometry that varies greatly between the rooms.Our experiments were performed on laser-range data fromthe COLD-Stockholm dataset [2]. The dataset contains multiple data sequences captured using a mobile robot navigatingwith constant speed through four different floors of anoffice building. On each floor, the robot navigates throughrooms of different semantic categories. There are 9 differentlarge offices, 8 different small offices (distributed acrossthe floors), 4 long corridors (1 per floor, with varyingappearance in different parts), and multiple examples ofplaces in doorways. The dataset features several other roomcategories: an elevator, a living room, a meeting room, a largemeeting room, and a kitchen. However, with only one ortwo room instances in each category. Therefore, we decidedto designate those categories as novel when testing noveltydetection and used the remaining four categories for themajority of the experiments. To ensure variability betweenthe training and test sets, we split the data samples four times,each time training the DGSM model on samples from threefloors and leaving one floor out for testing. The presentedresults are averaged over the four splits.The experiments were conducted using LibSPN [16]. SPNsare still a new architecture, and only few, limited domainspecific implementations exist at the time of writing. Incontrast, our library offers a general toolbox for structure generation, learning and inference, and enables quickapplication of SPNs to new domains. It integrates withTensorFlow, which leads to an efficient solution capable ofutilizing multiple GPUs, and enables combining SPNs withother deep architec

on widely used geometrical laser-range features [12][13]. Second, we benchmark novelty detection against one-class SVM trained on the same features. In both cases, DGSM offers superior accuracy. Finally, we compare the generative properties of our model to Generative Adversarial Networks (GANs) [14][15] on the two remaining inference tasks,