Learning Equality Constraints for Motion Planning on Manifolds
Giovanni Sutanto, Isabel M. Rayas Fernández, Peter Englert, Ragesh K. Ramachandran, Gaurav S. Sukhatme
LLearning Equality Constraints forMotion Planning on Manifolds
Giovanni Sutanto*, Isabel M. Rayas Fern´andez, Peter Englert,Ragesh K. Ramachandran, Gaurav S. Sukhatme
Robotic Embedded Systems LaboratoryUniversity of Southern California
Abstract:
Constrained robot motion planning is a widely used technique to solvecomplex robot tasks. We consider the problem of learning representations of con-straints from demonstrations with a deep neural network, which we call Equal-ity Constraint Manifold Neural Network (ECoMaNN). The key idea is to learna level-set function of the constraint suitable for integration into a constrainedsampling-based motion planner. Learning proceeds by aligning subspaces in thenetwork with subspaces of the data. We combine both learned constraints and ana-lytically described constraints into the planner and use a projection-based strategyto find valid points. We evaluate ECoMaNN on its representation capabilities ofconstraint manifolds, the impact of its individual loss terms, and the motions pro-duced when incorporated into a planner.
Video:
Code: https://github.com/gsutanto/smp_manifold_learning
Keywords: manifold learning, motion planning, learning from demonstration
Robots must be able to plan motions that follow various constraints in order for them to be usefulin real-world environments. Constraints such as holding an object, maintaining an orientation, orstaying within a certain distance of an object of interest are just some examples of possible restric-tions on a robot’s motion. In general, two approaches to many robotics problems can be described.One is the traditional approach of using handwritten models to capture environments, physics, andother aspects of the problem mathematically or analytically, and then solving or optimizing theseto find a solution. The other, popularized more recently, involves the use of machine learning toreplace, enhance, or simplify these hand-built parts. Both have challenges: Acquiring training datafor learning can be difficult and expensive, while describing precise models analytically can rangefrom tedious to impossible. Here, we approach the problem from a machine learning perspectiveand propose a solution to learn constraints from demonstrations. The learned constraints can be usedalongside analytical solutions within a motion planning framework.In this work, we propose a new learning-based method for describing motion constraints, calledEquality Constraint Manifold Neural Network (ECoMaNN). ECoMaNN learns a function whichevaluates a robot configuration on whether or not it meets the constraint, and for configurations nearthe constraint, on how far away it is. We train ECoMaNN with datasets consisting of configurationsthat adhere to constraints, and present results for kinematic robot tasks learned from demonstrations.We use a sequential motion planning framework to solve motion planning problems that are bothconstrained and sequential in nature, and incorporate the learned constraint representations into it.We evaluate the constraints learned by ECoMaNN with various datasets on their representationquality. Further, we investigate the usability of learned constraints in sequential motion planningproblems. * Giovanni Sutanto is now at X Development LLC. He contributed to this work during his past affiliation withthe Robotic Embedded Systems Laboratory at USC. a r X i v : . [ c s . R O ] S e p Related work
Manifold learning is applicable to many fields and thus there exist a wide variety of methods for it.Linear methods include PCA and LDA [1], and while they are simple, they lack the complexity torepresent complex manifolds. Nonlinear methods include multidimensional scaling (MDS), locallylinear embedding (LLE), Isomap, and local tangent space alignment (LTSA). These approaches usetechniques such as eigenvalue decomposition, nearest neighbor reconstructions, and local-structure-preserving graphs to visualize and represent manifolds. In LTSA, the local tangent space informationof each point is aligned to create a global representation of the manifold. We refer the reader to[2] for details. Recent work in manifold learning additionally takes advantage of (deep) neuralarchitectures. Some approaches use autoencoder-like models [3, 4] or deep neural networks [5]to learn manifolds, e.g. of human motion. Others use classical methods combined with neuralnetworks, for example as a loss function for control [6] or as structure for the network [7]. LocallySmooth Manifold Learning (LSML) [8] defines and learns a function which describes the tangentspace of the manifold, allowing randomly sampled points to be projected onto it. Our work is relatedto many of these approaches; in particular, the tangent space alignment in LTSA is an idea thatECoMaNN uses extensively. Similar to the ideas presented in this paper, the work in [9] delineatesan approach to solve motion planning problems by learning the solution manifold of an optimizationproblem. In contrast to others, our work focuses on learning implicit functions of equality constraintmanifolds, which is a generalization of the learning representations for Signed Distance Fields (SDF)[10, 11], up to a scale, for higher-dimensional manifolds.
Learning from demonstration (LfD) techniques learn a task representation from data which is usableto generate robot motions that imitate the demonstrated behavior. One approach to LfD is inverseoptimal control (IOC), which aims to find a cost function that describes the demonstrated behav-ior [12, 13, 14, 15]. Recently, IOC has been extended to extract constraints from demonstrations[16, 17]. There, a cost function as well as equality and inequality constraints are extracted fromdemonstrations, which are useful to describe behavior like contacts or collision avoidance. Ourwork can be seen as a special case where the task is only represented in form of constraints. Insteadof using the extracted constraints in optimal control methods, we integrate them into sampling-based motion planning methods, which are not parameterized by time and do not suffer from poorinitializations. A more direct approach to LfD is to learn parameterized motion representations[18, 19, 20]. They represent the demonstrations in a parameterized form such as Dynamic Move-ment Primitives [21]. Here, learning a primitive from demonstration is often possible via linearregression; however, the ability to generalize to new situations is more limited. Other approaches toLfD include task space learning [22] and deep learning [23]. We refer the reader to the survey [24]for a broad overview on LfD.
Sampling-based motion planning (SBMP) is a broad field which tackles the problem of motion plan-ning by using randomized sampling techniques to build a tree or graph of configurations (also calledsamples), which can then be used to plan paths between configurations. Many SBMP algorithmsderive from rapidly-exploring random trees (RRT) [25], probabilistic roadmaps (PRM) [26], or theiroptimal counterparts [27]. A more challenging and realistic motion planning task is that of con-strained SBMP [28], where there are motion constraints beyond just obstacle avoidance which leadto a free configuration space manifold of lower dimension than the ambient configuration space.Previous research has also investigated incorporating learned constraints or manifolds into planningframeworks. These include performing planning in learned latent spaces [29], learning a bettersampling distribution in order to take advantage of the structure of valid configurations rather thanblindly sample uniformly in the search space [30, 31], and attempting to approximate the manifold(both explicitly and implicitly) of valid points with graphs in order to plan on them more effec-tively [32, 33, 34]. Our method differs from previous work in that ECoMaNN learns an implicitdescription of a constraint manifold via a level set function, and during planning, we assume this2epresentation for each task. We note that our method could be combined with others, e.g. learnedsampling distributions, to further improve planning results.
Here we present the necessary background on manifold theory [35]. Informally, a manifold is a sur-face which can be well-approximated locally using an open set of a Euclidean space near every point.Manifolds are generally defined using charts , which are collections of open sets whose union yieldsthe manifold, and a coordinate map , which is a continuous map associated with each set. However,an alternative representation which is useful from a computational perspective is to represent themanifold as the zero level set of a continuous function. Since the latter representation is a directresult of the implicit function theorem, it is referred to as implicit representation of the manifold.For example, the manifold represented by the zero level set of the function h M ( x, y ) = x + y − (i.e. { ( x, y ) | x + y − } ) is a circle. Moreover, the implicit function associated with a smoothmanifold is smooth. Thus, we can associate a manifold with every equality constraint. The vectorspace containing the set of all tangent vectors at q is denoted using T q M . Given a manifold with thecorresponding implicit function h M ( q ) , if we endow its tangent spaces with an appropriate innerproduct structure, then such a manifold is often referred as a Riemannian manifold. In this work themanifolds are assumed to be Riemannian. In this work, we aim to integrate learned constraint manifolds into a motion planning framework[36]. The motion planner considers kinematic motion planning problems in a configuration space
C ⊆ R d . A robot configuration q ∈ C describes the state of one or more robots with d degreesof freedom in total. A manifold M is represented as an equality constraint h M ( q ) = . The setof robot configurations that are on a manifold M is given by C M = { q ∈ C | h M ( q ) = } . Theplanning problem is defined as a sequence of ( m + 1) such manifolds M = { M , M , . . . , M m +1 } and an initial configuration q start ∈ C M on the first manifold. The goal is to find a path from q start that traverses the manifold sequence M and reaches a configuration on the goal manifold M m +1 .A path on the i -th manifold is defined as τ i : [0 , → C M i and J ( τ i ) is the cost function of apath J : T → R ≥ where T is the set of all non-trivial paths. The problem is formulated as anoptimization over a set of paths τ = ( τ , . . . , τ m ) that minimizes the sum of path costs under theconstraints of traversing M : τ (cid:63) = arg min τ m (cid:88) i =1 J ( τ i ) s.t. τ (0) = q start , τ m (1) ∈ C M m +1 ∩ C free ,m +1 , τ i (1) = τ i +1 (0) ∀ i =1 ,...,m − C free ,i +1 = Υ( C free ,i , τ i ) ∀ i =1 ,...,m , τ i ( s ) ∈ C M i ∩ C free ,i ∀ i =1 ,...,m ∀ s ∈ [0 , (1) Υ is an operator that describes the change in the free configuration space (the space of all configu-rations that are not in collision with the environment) C free when transitioning to the next manifold.The operator Υ is not explicitly known and we only assume to have access to a collision checker thatdepends on the current robot configuration and the object locations in the environment. Intelligentlyperforming goal-achieving manipulations that change the free configuration space forms a key chal-lenge in robot manipulation planning. The SMP ∗ (Sequential Manifold Planning) algorithm is ableto solve this problem for a certain class of motion planning scenarios. It iteratively applies RRT ∗ tofind a path that reaches the next manifold while staying on the current manifold. For further detailsof the SMP ∗ algorithm, we refer the reader to [36]. In this paper, we employ data-driven learningmethods to learn individual equality constraints h M ( q ) = 0 from demonstrations with the goal tointegrate them with analytically defined manifolds into this framework. We propose a novel neural network structure, called
Equality Constraint Manifold Neural Net-work (ECoMaNN), which is a (global) equality constraint manifold learning representation that3nforces the alignment of the (local) tangent spaces and normal spaces with the information ex-tracted from the Local Principal Component Analysis (Local PCA) [37] of the data. ECoMaNNtakes a configuration q as input and outputs the prediction of the implicit function h M ( q ) . Wetrain ECoMaNN in a supervised manner, from demonstrations. One of the challenges is that thesupervised training dataset is collected only from demonstrations of data points which are on themanifold C M , called the on-manifold dataset. Collecting both the on-manifold C M and off-manifold C \ M = { q ∈ C | h M ( q ) (cid:54) = } datasets for supervised training would be tedious because the im-plicit function h M values of points in C \ M are typically unknown and hard to label. We show that,though our approach is only given data on C M , it can still learn a useful and sufficient representationof the manifold for use in planning. Figure 1: A visualiza-tion of data augmenta-tion along the 1D nor-mal space of a point q in 3D space. Here,purple points are thedataset, pink points arethe K NN of q , andthe dark red point is ˇ q . q is at the axes ori-gin, and the green planeis the approximated tan-gent space at that point.Our goal is to learn a single global representation of the constraint man-ifold M in the form of a neural network. Our approach leverages lo-cal information on the manifold in the form of the tangent and normalspaces [38]. With regard to h M , the tangent and normal spaces areequivalent to the null and row space, respectively, of the Jacobian ma-trix J M (´ q ) = ∂h M ( q ) ∂ q (cid:12)(cid:12)(cid:12) q =´ q , and valid in a small neighborhood aroundthe point ´ q . Using on-manifold data, the local information of the man-ifold can be analyzed using Local PCA. For each data point q in theon-manifold dataset, we establish a local neighborhood using K -nearestneighbors ( K NN) ˆ K = { ˆ q , ˆ q , . . . , ˆ q K } , with K ≥ d . After a changeof coordinates, q becomes the origin of a new local coordinate frame F L ,and the K NN becomes ˜ K = { ˜ q , ˜ q , . . . , ˜ q K } with ˜ q k = ˆ q k − q for allvalues of k . Defining the matrix X = [˜ q ˜ q . . . ˜ q K ] T ∈ R K × d ,we can compute the covariance matrix S = K − X T X ∈ R d × d . Theeigendecomposition of S = VΣV T gives us the Local PCA. The ma-trix V contains the eigenvectors of S as its columns in decreasing orderw.r.t. the corresponding eigenvalues in the diagonal matrix Σ . Theseeigenvectors form the basis of the coordinate frame F L .This local coordinate frame F L is tightly related to the tangent space T q M and the normal space N q M of the manifold M at q . That is, the ( d − l ) eigenvectors corresponding to the ( d − l ) biggest eigenvalues of Σ form a basis of T q M , whilethe remaining l eigenvectors form the basis of N q M . Furthermore, the l smallest eigenvalues of Σ will be close to zero, resulting in the l eigenvectors associated with them forming the basis of thenull space of S . On the other hand, the remaining ( d − l ) eigenvectors form the basis of the row spaceof S . We follow the same technique as Deutsch and Medioni [38] for automatically determining thenumber of constraints l from data, which is also the number of outputs of ECoMaNN . Suppose theeigenvalues of S are { λ , λ , . . . , λ d } (in decreasing order w.r.t. magnitude). Then the number ofconstraints can be determined as l = arg max ([ λ − λ , λ − λ , . . . , λ d − − λ d ]) .We now present several methods to define and train ECoMaNN, as follows: ECoMaNN aims to align the following:(a) the null space of J M and the row space of S (which must be equivalent to tangent space T q M )(b) the row space of J M and the null space of S (which must be equivalent to normal space N q M )for the local neighborhood of each point q in the on-manifold dataset. Suppose the eigenvectors of S are { v , v , . . . , v d } and the singular vectors of J M are { e , e , . . . , e d } , where the indices indicatedecreasing order w.r.t. the eigenvalue/singular value magnitude. The null spaces of S and J M arespanned by { v d − l +1 , . . . , v d } and { e l +1 , . . . , e d } , respectively. The two conditions above implythat the projection of the null space eigenvectors of J M into the null space of S should be , andsimilarly for the row spaces. Hence, we achieve this by training ECoMaNN to minimize projectionerrors (cid:13)(cid:13) V N V TN E N (cid:13)(cid:13) and (cid:13)(cid:13) E N E TN V N (cid:13)(cid:13) with V N = [ v d − l +1 . . . v d ] and E N = [ e l +1 . . . e d ] at all the points in the on-manifold dataset. Here we assume that the intrinsic dimensionality of the manifold at each point remains constant. .2 Data augmentation with off-manifold data The training dataset is on-manifold, i.e., each point q in the dataset satisfies h M ( q ) = . ThroughLocal PCA on each of these points, we know the data-driven approximation of the normal spaceof the manifold at q . Hence, we know the directions where the violation of the equality constraintincreases, i.e., the same direction as any vector picked from the approximate normal space. Since ourfuture use of the learned constraint manifold on motion planning does not require the acquisition ofthe near-ground-truth value of h M ( q ) (cid:54) = , we can set this off-manifold valuation of h M arbitrarily,as long as it does not interfere with the utility for projecting an off-manifold point onto the manifold.Therefore, we can augment our dataset with off-manifold data to achieve a more robust learning ofECoMaNN. For each point q in the on-manifold dataset, and for each random unit vector u pickedfrom the normal space at q , we can add an off-manifold point ˇ q = q + i(cid:15) u with a positive integer i and a small positive scalar (cid:15) (see Figure 1 for a visualization). However, if the closest on-manifolddata point to an augmented point ˇ q = q + i(cid:15) u is not q , we reject it. This prevents situations likeaugmenting a point on a sphere beyond the center of the sphere. Data augmentation is a techniqueused in various fields, and our approach has similarities to others [39, 40, 41], though in this workwe focus on using augmentation to aid learning an implicit constraint function for robotic motionplanning. With this data augmentation, we now define several losses used to train ECoMaNN. For the augmented data point ˇ q = q + i(cid:15) u , we set the label satisfying (cid:107) h M (ˇ q ) (cid:107) = i(cid:15) . During training, we minimize thenorm prediction error L norm = (cid:107) ( (cid:107) h M (ˇ q ) (cid:107) − i(cid:15) ) (cid:107) for each augmented point ˇ q .Furthermore, we define the following three siamese losses. The main intuition behind these lossesis that we expect the learned function h M to output similar values for similar points. Siamese loss for reflection pairs
For the augmented data point ˇ q = q + i(cid:15) u and its reflection pair q − i(cid:15) u , we can expect that h M ( q + i(cid:15) u ) = − h M ( q − i(cid:15) u ) , or in other words, that an augmentedpoint and its reflection pair should have the same h M but with opposite signs. Therefore, duringtraining we minimize the siamese loss L reflection = (cid:107) h M ( q + i(cid:15) u ) + h M ( q − i(cid:15) u ) (cid:107) . Siamese loss for augmentation fraction pairs
Similarly, between the pair ˇ q = q + i(cid:15) u and q + ab i(cid:15) u , where a and b are positive integers satisfying < ab < , we can expectthat h M ( q + i(cid:15) u ) (cid:107) h M ( q + i(cid:15) u ) (cid:107) = h M ( q + ab i(cid:15) u ) (cid:107) h M ( q + ab i(cid:15) u ) (cid:107) . In other words, the normalized h M values should bethe same on an augmented point q + i(cid:15) u as on any point in between the on-manifold point q and that augmented point q + i(cid:15) u . Hence, during training we minimize the siamese loss L fraction = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) h M ( q + i(cid:15) u ) (cid:107) h M ( q + i(cid:15) u ) (cid:107) − h M ( q + ab i(cid:15) u ) (cid:107) h M ( q + ab i(cid:15) u ) (cid:107) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . Siamese loss for similar augmentation pairs
Suppose for nearby on-manifold data points q a and q c , their approximate normal spaces N q a M and N q c M are spanned by eigenvector bases F aN = { v ad − l +1 , . . . , v ad } and F cN = { v cd − l +1 , . . . , v cd } , respectively. If F aN and F cN are closely aligned,the random unit vectors u a from F aN and u c from F cN can be obtained by u a = (cid:80) dj = d − l +1 w j v aj (cid:107) (cid:80) dj = d − l +1 w j v aj (cid:107) and u c = (cid:80) dj = d − l +1 w j v cj (cid:107) (cid:80) dj = d − l +1 w j v cj (cid:107) , where { w d − l +1 , . . . , w d } are random scalar weights from a standardnormal distribution common to both the bases of F aN and F cN . This will ensure that u a and u c arealigned as well, and we can expect that h M ( q a + i(cid:15) u a ) = h M ( q c + i(cid:15) u c ) . In other words, twoaligned augmented points in the same level set should have the same h M value. Therefore, duringtraining we minimize the siamese loss L similar = (cid:107) h M ( q a + i(cid:15) u a ) − h M ( q c + i(cid:15) u c ) (cid:107) . In general,the alignment of F aN and F cN is not guaranteed, for example due to the numerical sensitivity ofsingular value/eigen decomposition. Therefore, we introduce an algorithm for Orthogonal SubspaceAlignment (OSA) in the Supplementary Material to ensure that this assumption is satisfied.While L norm governs only the norm of ECoMaNN’s output, the other three losses L reflection , L fraction ,and L similar constrain the (vector) outputs of ECoMaNN based on pairwise input data points withoutexplicitly hand-coding the desired output itself. We avoid the hand-coding of the desired outputbecause this is difficult for high-dimensional manifolds, except when there is prior knowledge aboutthe manifold available, such as in the case of Signed Distance Fields (SDF) manifolds.5 a) Samples from ECoMaNN (b) Samples from VAE P r o j e c t i o n s u cc e ss [ % ] (c) ECoMaNN projection success Figure 2: Images a and b visualize a slice near z = 0 of the Plane dataset for experiment 5.1. Redpoints are the training dataset and blue points are samples generated from the learned manifolds.The points projected onto the manifold using ECoMaNN are closer to the manifold, with an 85%projection success rate. A significant portion of the points generated using the VAE lie inside thesurface, which leads to a lower success rate of 77%. Figure c shows the projection success ofECoMaNN over the number of training iterations. The quantitative results are found in Table 1.Table 1: Accuracy and precision of learned manifolds averaged across 3 instances. “Train” indi-cates results on the on-manifold training set; “test” indicates N = 1000 projected (ECoMaNN) orsampled (VAE) points. ECoMaNN VAEDataset µ ¯ h M (train) µ ¯ h M (test) P ¯ h M µ ¯ h M (train) µ ¯ h M (test) P ¯ h M Sphere . ± .
009 0 . ± .
009 100 . ± . . ± .
088 0 . ± .
165 46 . ± .
3D Circle . ± .
011 0 . ± .
011 78 . ± . . ± .
074 0 . ± .
069 0 . ± . Plane . ± .
005 0 . ± .
005 88 . ± . . ± .
075 0 . ± .
216 77 . ± . Orient . ± .
009 0 . ± .
009 73 . ± . . ± .
037 0 . ± .
237 85 . ± . We use the robot simulator MuJoCo [42] to generate four datasets. The size of each dataset isdenoted as N . We define a ground truth constraint ¯ h M , randomly sample points in the configuration(joint) space, and use a constrained motion planner to find robot configurations in C M that producethe on-manifold datasets: Sphere : 3D point that has to stay on the surface of a sphere. N =5000 , d = 3 , l = 1 .
3D Circle : A point that has to stay on a circle in 3D space. N = 1000 , d =3 , l = 2 . Plane : Robot arm with 3 rotational DoFs where the end effector has to be on a plane. N = 20000 , d = 3 , l = 1 . Orient : Robot arm with 6 rotational DoFs that has to keep its orientationupright (e.g., transporting a cup). N = 21153 , d = 6 , l = 2 . In the following experiments, weparametrize ECoMaNN with hidden layers of size , , , and . The hidden layers use a tanh activation function and the output layer is linear. We compare the accuracy and precision of the manifolds learned by ECoMaNN with those learnedby a variational autoencoder (VAE) [43]. VAEs are a popular generative model that embeds datapoints as a distribution in a learned latent space, and as such new latent vectors can be sampled anddecoded into new examples which fit the distribution of the training data. We use two metrics: First,the distance µ ¯ h M which measures how far a point is away from the ground-truth manifold ¯ h M andwhich we evaluate for both the training data points and randomly sampled points, and second, thepercent P ¯ h M of randomly sampled points that are on the manifold ¯ h M . We use a distance thresholdof . to determine success when calculating P ¯ h M . For ECoMaNN, randomly sampled points areprojected using gradient descent with the learned implicit function until convergence. For the VAE,latent points are sampled from N (0 , and decoded into new configurations. We also tested classical manifold learning techniques (Isomap, LTSA, PCA, MDS, and LLE). We foundthem empirically not expressive enough and/or unable to support projection or sampling operations, necessarycapabilities for this work. ± ± ± ± ± ± ± ± ± ± ± ± L reflection (92.67 ± ± ± L fraction (88.33 ± ± ± L similar (83.00 ± ± ± (cid:15) to the square root of the mean eigenvalues of theapproximate tangent space, which we found to work well experimentally. With the exceptions of theembedding size and the input size, which are set to the dimensionality d − l as the tangent space ofthe constraint learned by ECoMaNN and the ambient space dimensionality d of the dataset, respec-tively, the VAE has the same parameters for each dataset: 4 hidden layers with 128, 64, 32, and 16units in the encoder and the same but reversed in the decoder; the weight of the KL divergence loss β = 0 . ; using batch normalization; and trained for 100 epochs.Our results show that for every dataset except Orient, ECoMaNN out performs the VAE in bothmetrics. ECoMaNN additionally outperforms the VAE with the Orient dataset in the testing phase,which suggests more robustness of the learned model. We find that though the VAE also performsrelatively well in most cases, it cannot learn a good representation of the 3D Circle constraint andfails to produce any valid sampled points. ECoMaNN, on the other hand, can learn to represent allfour constraints well. In the ablation study, we compare P ¯ h M across 7 different ECoMaNN setups: 1) no ablation; 2) with-out data augmentation; 3) without orthogonal subspace alignment (OSA) during data augmentation;4) without siamese losses during training; 5) without L reflection ; 6) without L fraction ; and 7) without L similar . Results are reported in Table 2. The data suggest that all parts of the training process areessential for a high success rate during projection. Of the features tested, data augmentation ap-pears to have the most impact. This makes sense because without augmented data to train on, anyconfiguration that does not already lie on the manifold will have undefined output when evaluatedwith ECoMaNN. Additionally, results from ablating the individual siamese losses suggest that thecontribution of each is dependent on the context and structure of the constraint. Complementary tothis ablation study, we present some additional experimental results in the Supplementary Material.7igure 5: The images visualize a path that was planned on a learned orientation manifold. Figure 4: Planned path onthe learned manifold (red)and on the ground truthmanifold (black).In the final experiment, we integrate ECoMaNN into the sequential mo-tion planning framework described in Section 3.2. We mix the learnedconstraints with analytically defined constraints and evaluate it for twotasks. The first one is a geometric task, visualized in Figure 4, where apoint starting on a paraboloid in 3D space must find a path to a goal stateon another paraboloid. The paraboloids are connected by a sphere, andthe point is constrained to stay on the surfaces at all times. In this case,we give the paraboloids analytically to the planner, and use ECoMaNNto learn the connecting constraint using the Sphere dataset. Figure 4shows the resulting path where the sphere is represented by a learnedmanifold (red line) and where it is represented by the ground-truth man-ifold (black line). While the paths do not match exactly, both paths areon the manifold and lead to similar solutions in terms of path lengths.The task was solved in .
09 s on a 2.2 GHz Intel Core i7 processor.The tree explored nodes and the found path consists of nodes.The second task is a robot pick-and-place task with the additional con-straint that the transported object needs to be oriented upwards through-out the whole motion. For this, we use the Orient dataset to learnthe manifold for the transport phase and combine it with other man-ifolds that describe the pick and place operation. The planning timewas .
97 s , the tree contained nodes and the optimal path had nodes. Images of the resulting path are shown in Figure 5. In this paper, we presented a novel method called Equality Constraint Manifold Neural Networkfor learning equality constraint manifolds from data. ECoMaNN works by aligning the row andnull spaces of the local PCA and network Jacobian, which results in approximate learned normaland tangent spaces of the underlying manifold, suitable for use within a constrained sampling-basedmotion planner. In addition, we introduced a method for augmenting a purely on-manifold dataset toinclude off-manifold points and several loss functions for training. This improves the robustness ofthe learned method while avoiding hand-coding the labels for the augmented points. We also showedthat the learned manifolds can be used in a sequential motion planning framework for constrainedrobot tasks.While our experiments show success in learning a variety of manifolds, there are some limitations toour method. First, ECoMaNN by design can only learn equality constraints. Although many taskscan be specified with such constraints, inequality constraints are also an important part of manyrobot planning problems. Additionally, because of inherent limitations in learning from data, ECo-MaNN does not guarantee that a randomly sampled point in configuration space will be projectedsuccessfully onto the learned manifold. This presents challenges in designing asymptotically com-plete motion planning algorithms, and is an important area of research. In the future, we plan onfurther testing ECoMaNN on more complex tasks, and in particular on tasks which are demonstratedby a human rather than from simulation. 8 cknowledgments
This material is based upon work supported by the National Science Foundation Graduate ResearchFellowship Program under Grant No. DGE-1842487. Any opinions, findings, and conclusions orrecommendations expressed in this material are those of the author(s) and do not necessarily reflectthe views of the National Science Foundation. This work was supported in part by the Office ofNaval Research (ONR) under grant N000141512550.
References [1] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern classification.
John Wiley & Sons, Inc. , 2001.[2] Y. Ma and Y. Fu.
Manifold learning theory and applications . CRC press, 2011.[3] D. Holden, J. Saito, T. Komura, and T. Joyce. Learning motion manifolds with convolutionalautoencoders. In
ACM SIGGRAPH Asia 2015 Technical Briefs , 2015.[4] N. Chen, M. Karl, and P. van der Smagt. Dynamic movement primitives in latent space of time-dependent variational autoencoders. In
IEEE-RAS 16th Humanoids , pages 629–636, 2016.[5] X. S. Nguyen, L. Brun, O. L´ezoray, and S. Bougleux. A neural network based on spd manifoldlearning for skeleton-based hand gesture recognition. In
IEEE CVPR , 2019.[6] G. Sutanto, N. Ratliff, B. Sundaralingam, Y. Chebotar, Z. Su, A. Handa, and D. Fox. Learninglatent space dynamics for tactile servoing. In
IEEE ICRA , pages 3622–3628, 2019.[7] W. Wang, Y. Huang, Y. Wang, and L. Wang. Generalized autoencoder: A neural networkframework for dimensionality reduction. In
IEEE CVPR Workshops , pages 490–497, 2014.[8] P. Doll´ar, V. Rabaud, and S. Belongie. Non-isometric manifold learning: Analysis and analgorithm. In , pages 241–248, 2007.[9] T. Osa. Learning the solution manifold in optimization and its application in motion planning. arXiv , 2020. URL https://arxiv.org/pdf/2007.12397.pdf .[10] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning contin-uous signed distance functions for shape representation. In
IEEE CVPR , 2019.[11] J. Mahler, S. Patil, B. Kehoe, J. Van Den Berg, M. Ciocarlie, P. Abbeel, and K. Goldberg.Gp-gpis-opt: Grasp planning with shape uncertainty using gaussian process implicit surfacesand sequential convex programming. In
IEEE ICRA , pages 4919–4926, 2015.[12] N. D. Ratliff, J. A. Bagnell, and M. Zinkevich. Maximum margin planning. In
ICML , 2006.[13] B. D. Ziebart, A. Maas, J. A. Bagnell, and A. K. Dey. Maximum Entropy Inverse Reinforce-ment Learning.
In Proceedings of the AAAI Conference on Artificial Intelligence , 2008.[14] N. D. Ratliff, D. Silver, and J. A. Bagnell. Learning to search: Functional gradient techniquesfor imitation learning.
Autonomous Robots , 27:25–53, 2009.[15] S. Levine and V. Koltun. Continuous inverse Optimal Control with Locally Optimal Examples.In
ICML , 2012.[16] A.-S. Puydupin-Jamin, M. Johnson, and T. Bretl. A convex approach to inverse optimal controland its application to modeling human locomotion. In
IEEE ICRA , 2012.[17] P. Englert, N. A. Vien, and M. Toussaint. Inverse kkt — learning cost functions of manipulationtasks from demonstrations.
IJRR , 36(13-14):1474–1488, 2017.[18] S. Schaal, A. Ijspeert, and A. Billard. Computational Approaches to Motor Learning by Imi-tation.
Philosophical Transactions of the Royal Society of London , 358:537–547, 2003.[19] A. Paraschos, C. Daniel, J. Peters, and G. Neumann. Probabilistic Movement Primitives. In
NIPS , 2013. 920] P. Pastor, M. Kalakrishnan, S. Chitta, E. Theodorou, and S. Schaal. Skill learning and taskoutcome prediction for manipulation. In
IEEE ICRA , 2011.[21] A. J. Ijspeert, J. Nakanishi, H. Hoffmann, P. Pastor, and S. Schaal. Dynamical movementprimitives: Learning attractor models for motor behaviors.
Neural Computation , 25(2), 2013.[22] N. Jetchev and M. Toussaint. TRIC: Task space retrieval using inverse optimal control.
Au-tonomous Robots , 37(2):169–189, 2014.[23] C. Finn, S. Levine, and P. Abbeel. Guided cost learning: Deep inverse optimal control viapolicy optimization. In
ICML , 2016.[24] B. D. Argall, S. Chernova, M. Veloso, and B. Browning. A survey of robot learning fromdemonstration.
Robotics and Autonomous Systems , 57(5):469–483, 2009.[25] S. M. LaValle. Rapidly-exploring random trees: A new tool for path planning. TechnicalReport TR 98-11, Computer Science Department, Iowa State University, 1998.[26] L. Kavraki and J.-C. Latombe. Randomized preprocessing of configuration for fast path plan-ning. In
IEEE ICRA , 1994.[27] S. Karaman and E. Frazzoli. Sampling-based algorithms for optimal motion planning.
Inter-national Journal of Robotics Research , 30:846–894, 2011.[28] Z. Kingston, M. Moll, and L. E. Kavraki. Sampling-based methods for motion planning withconstraints.
Annual review of control, robotics, and autonomous systems , 1:159–185, 2018.[29] B. Ichter and M. Pavone. Robot motion planning in learned latent spaces.
IEEE Robotics andAutomation Letters , 4(3):2407–2414, 2019.[30] B. Ichter, J. Harrison, and M. Pavone. Learning sampling distributions for robot motion plan-ning. In
IEEE ICRA , 2018.[31] R. Madaan, S. Zeng, B. Okorn, and S. Scherer. Learning adaptive sampling distributions formotion planning by self-imitation.
Workshop on Machine Learning in Robot Motion Planning,IEEE IROS , 2018.[32] M. Phillips, B. Cohen, S. Chitta, and M. Likhachev. E-graphs: Bootstrapping planning withexperience graphs.
Robotics: Science and Systems , 2012.[33] I. Havoutis and S. Ramamoorthy. Motion synthesis through randomized exploration on sub-manifolds of configuration space. In
Robot Soccer World Cup , pages 92–103. Springer, 2009.[34] F. Zha, Y. Liu, W. Guo, P. Wang, M. Li, X. Wang, and J. Li. Learning the metric of taskconstraint manifolds for constrained motion planning.
Electronics , 7(12):395, 2018.[35] W. M. Boothby.
An introduction to differentiable manifolds and Riemannian geometry; 2nded.
Pure Appl. Math. Academic Press, Orlando, FL, 1986.[36] P. Englert, I. M. R. Fern´andez, R. K. Ramachandran, and G. S. Sukhatme. Sampling-BasedMotion Planning on Manifold Sequences. arXiv:2006.02027, 2020.[37] N. Kambhatla and T. K. Leen. Dimension Reduction by Local Principal Component Analysis.
Neural Comput. , 9(7):14931516, Oct. 1997.[38] S. Deutsch and G. Medioni. Unsupervised learning using the tensor voting graph. In
ScaleSpace and Variational Methods in Computer Vision , pages 282–293, 2015.[39] T.-J. Chin and D. Suter. Out-of-sample extrapolation of learned manifolds.
IEEE Transactionson Pattern Analysis and Machine Intelligence , 30(9):1547–1556, 2008.[40] C. Bellinger, C. Drummond, and N. Japkowicz. Manifold-based synthetic oversampling withmanifold conformance estimation.
Machine Learning , 107(3):605–637, 2018.[41] K. Patel, W. Beluch, D. Zhang, M. Pfeiffer, and B. Yang. On-manifold adversarial data aug-mentation improves uncertainty calibration. arXiv preprint arXiv:1912.07458 , 2019.1042] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In
IEEE IROS , 2012.[43] D. P. Kingma and M. Welling. Auto-Encoding Variational Bayes.
CoRR , abs/1312.6114, 2014.[44] K. Thopalli, R. Anirudh, J. J. Thiagarajan, and P. Turaga. Multiple subspace alignment im-proves domain adaptation. In
ICASSP 2019 - 2019 IEEE International Conference on Acous-tics, Speech and Signal Processing (ICASSP) , pages 3552–3556, 2019.[45] X. He. Quantum subspace alignment for domain adaptation. arXiv:2001.02472, 2020.[46] A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison,L. Antiga, and A. Lerer. Automatic differentiation in pytorch. In
NIPS-W , 2017.[47] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. Corrado, A. Davis, J. Dean,M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefow-icz, L. Kaiser, M. Kudlur, J. Levenberg, D. Man, R. Monga, S. Moore, D. Murray, C. Olah,M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Va-sudevan, F. Vigas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng.Tensorflow: Large-scale machine learning on heterogeneous distributed systems, 2015. URL http://download.tensorflow.org/paper/whitepaper2015.pdf .11
Supplementary Materials
We provide the following supplementary material to enhance the main paper: • Orthogonal Subspace Alignment – In Section A.1, we describe the details of the orthogo-nal subspace alignment and provide an algorithm that shows its step-by-step computations. • Additional Experiments – In order to thoroughly evaluate the components of our ap-proach, we present some additional experimental results in Section A.2.
A.1 Orthogonal Subspace Alignment (OSA)
In previous work [44, 45], subspace alignment techniques – without orthogonality constraints – havebeen introduced to improve domain adaptation. For the purposes of this paper, we require a subspacealignment algorithm that preserves orthogonality of the subspaces being aligned, which we presentin this section.Given a set of orthonormal vectors F = { b , b , . . . , b d } which spans a space , the matrix B = [ b b . . . b d ] ∈ R d × d belongs to the Orthogonal Group O ( d ) . The Orthogonal Grouphas two connected components, where one connected component called the Special OrthogonalGroup SO ( d ) is characterized by determinant , and the other is characterized by determinant − .However, if B = [ b b . . . b d ] has determinant 1 (i.e. if B ∈ SO ( d ) ), then substituting b with its additive inverse ( − b ) will result in ¯ B = [ − b b . . . b d ] with determinant − . Align-ing two coordinate frames F a and F c to have a common origin and associated basis matrices B a and B c , respectively, is equivalent to finding an R ∈ SO ( d ) such that B a R = B c . The solutionto this problem exists if and only if B a and B c come from the same connected component of O ( d ) ,i.e. if either both B a , B c ∈ SO ( d ) or both determinants of B a and B c are − .For a subspace such as the normal space N q M associated with an on-manifold data point q on M spanned by the eigenvectors F N = { v d − l +1 , . . . , v d } , the concept of a determinant does notapply to V N = [ v d − l +1 . . . v d ] ∈ R d × l , as it is not a square matrix. However, the nor-mal space N q M can be described with infinitely-many orthonormal bases V N , V N , V N , ... V N ∞ where the set of column vectors of each is an orthonormal basis of N q M . Each of theseis a member of R d × l . Moreover, we can pick the transpose of one of them, for example V NT ,as a projection matrix, and V N as the inverse projection matrix. Applying the projection oper-ation to each of the orthonormal bases, we get W N = V NT V N = I l × l , W N = V NT V N , W N = V NT V N , ... W N ∞ = V NT V N ∞ , and we will show that W N , W N , W N , ..., W N ∞ are members of O ( l ) , which also has two connected components like O ( d ) . To showthis, first note that although V NT V N = I l × l , the matrix V N V NT (cid:54) = I d × d . Hence, for anymatrix A ∈ R d × d in general, V N V NT A (cid:54) = A . However, we will show that V N V NT v = v for any vector v in the vector space N q M . Suppose V N = [ b b . . . b l ] ∈ R d × l , thenwe can write V N V NT = (cid:80) li =1 b i b T i . Since the collection { b , b , . . . , b l } spans the vectorspace N q M , any vector v in this vector space can be expressed as v = (cid:80) li =1 α i b i . Moreover, b T i v = b T i (cid:80) lj =1 α j b j = α i for any i = 1 , , ..., l , because by definition of orthonormality b T i b j = 1 for i = j and b T i b j = 0 for i (cid:54) = j . Hence, V N V NT v = ( (cid:80) li =1 b i b T i ) v = (cid:80) li =1 ( b T i v ) b i = (cid:80) li =1 α i b i = v . Similarly, because the column vectors of V N , V N , V N , ..., V N ∞ are allinside the vector space N q M , it follows that V N V NT V N = V N , V N V NT V N = V N , V N V NT V N = V N , ..., V N V NT V N ∞ = V N ∞ . Similarly, it can be shown that V N i V NT i v = v for any vector v in the vector space N q M for any i = 0 , , , ..., ∞ . Furthermore, W NT W N = V NT ( V N V NT V N ) = V NT V N = I l × l , W NT W N = V NT ( V N V NT V N ) = V NT V N = I l × l , ... W NT ∞ W N ∞ = V NT ∞ ( V N V NT V N ∞ ) = V NT ∞ V N ∞ = I l × l , and W N W NT = V NT ( V N V NT V N ) = V NT V N = I l × l , W N W NT = V NT ( V N V NT V N ) = V NT V N = I l × l ,... W N ∞ W NT ∞ = V NT ( V N ∞ V NT ∞ V N ) = V NT V N = I l × l . All these show that W N , W N , W N , ..., W N ∞ ∈ O ( l ) . Moreover, using V N as the inverse projection matrix, we get V N = V N W N , V N = V N W N , V N = V N W N , ... V N ∞ = V N W N ∞ . Therefore,there is a one-to-one mapping between V N , V N , V N , ..., V N ∞ and W N , W N , W N , ..., W N ∞ . Furthermore, between any two of V N , V N , V N , ..., V N ∞ , e.g. V N i and V N j , there12xists R N ∈ SO ( l ) such that V N i R N = V N j if their SO ( l ) projections W N i and W N j both aremembers of the same connected component of O ( l ) .Now, suppose for nearby on-manifold data points q a and q c , their approximate normal spaces N q a M and N q c M are spanned by eigenvector bases F aN = { v ad − l +1 , . . . , v ad } and F cN = { v cd − l +1 , . . . , v cd } , respectively. Due to the curvature on the manifold M , the normal spaces N q a M and N q c M may intersect, but in general are different subspaces of R d × d . For the purpose of align-ing the basis of N q a M to the basis of N q c M , one may think to do projection of the basis vectorsof N q a M into N q c M . Problematically, this projection may result in a non-orthogonal basis of N q c M . Hence, we resort to an iterative method using a differentiable Special Orthogonal Group SO ( l ) . In particular, we form an l × l skew-symmetric matrix L ∈ so ( l ) with l ( l − / differ-entiable parameters –where so ( l ) is the Lie algebra of SO ( l ) , i.e. the set of all skew-symmetric l × l matrices–, and transform it through a differentiable exponential mapping (or matrix expo-nential) to get R N = exp( L ) with exp : so ( l ) → SO ( l ) . With V a N = (cid:2) v ad − l +1 . . . v ad (cid:3) and V c N = (cid:2) v cd − l +1 . . . v cd (cid:3) , we can do an iterative training process to minimize the alignment errorbetween V a N R N and V c N , that is L osa = (cid:13)(cid:13) I l × l − ( V a N R N ) T V c N (cid:13)(cid:13) . Depending on whether both W a N and W c N (which are the projections of V a N and V c N , respectively, to O ( l ) ) are members of the sameconnected component of O ( l ) or not, this alignment process may succeed or fail. However, if wedefine ¯ V a N = (cid:2) − v ad − l +1 v ad − l +2 . . . v ad (cid:3) and ¯ V c N = (cid:2) − v cd − l +1 v cd − l +2 . . . v cd (cid:3) , two outof the four pairs ( V a N , V c N ) , ( ¯ V a N , V c N ) , ( V a N , ¯ V c N ) , and ( ¯ V a N , ¯ V c N ) will be pairs in the same con-nected component. Thus, two of these pairs will achieve minimum alignment errors after trainingthe differentiable Special Orthogonal Groups SO ( l ) on these pairs, indicating successful alignment.These are the main insights for our local alignment of neighboring normal spaces of on-manifolddata points.For the global alignment of the normal spaces, we represent the on-manifold data points as a graph.Our Orthogonal Subspace Alignment (OSA) is outlined in Algorithm 1. We begin by constructinga sparse graph of nearest neighbor connections of each on-manifold data point, followed by theconstruction of this graph into an (un-directed) minimum spanning tree (MST), and eventually theconversion of the MST to a directed acyclic graph (DAG). This graph construction is detailed inlines 2 - 8 of Algorithm 1.Each directed edge in the DAG represents a pair of on-manifold data points whose normal spacesare to be aligned locally. Our insights for the local alignment of neighboring normal spaces areimplemented in lines 9 - 21 of Algorithm 1. In the actual implementation, these local alignmentcomputations are done as a vectorized computation which is faster than doing it in a for-loop aspresented in Algorithm 1; this for-loop presentation is made only for the sake of clarity. We initializethe l ( l − / differentiable parameters of the l × l skew-symmetric matrix L with near zero randomnumbers, which essentially will map to a near identity matrix I l × l of R N via the exp() mapping, asstated in line 14 of Algorithm 1 . This is reasonable because we assume that most of the neighboringnormal spaces are already/close to being aligned initially. We optimize the alignment of the fourpairs ( V a N , V c N ) , ( ¯ V a N , V c N ) , ( V a N , ¯ V c N ) , and ( ¯ V a N , ¯ V c N ) in lines 16 - 19 of Algorithm 1.Once the local alignments are done, the algorithm then traverses the DAG in breadth-first order,starting from the root r , where the orientation of the root is already chosen and committed to.During the breadth-first traversal of the DAG, three things are done: First, the orientation of eachpoint is chosen based on the minimum alignment loss; second, the local alignment transforms arecompounded/integrated along the path from root to the point; and finally, the (globally) alignedorthogonal basis of each point is computed and returned as the result of the algorithm. These stepsare represented by lines 22 – 48 of Algorithm 1. Although most of our ECoMaNN implementation is done in PyTorch [46], the OSA algorithm is imple-mented in TensorFlow [47], because at the time of implementation of the OSA algorithm, PyTorch did notsupport the differentiable matrix exponential (i.e. the exponential mapping) computation yet while TensorFlowdid. lgorithm 1 Orthogonal Subspace Alignment (OSA) function OSA( { ( q ∈ C M , orthogonal basis stacked as matrix V N associated with N q M ) } ) q ∈ C M with its H nearest neighbors, E ; H needs to be chosen to be a value as small as possible that { q ∈ C M \{ q r }} being reachable from the root point q r : G ← computeNearestNeighborsSparseGraph ( { q ∈ C M } , H ) T ← computeMinimumSpanningTree ( G ) E ← computeDirectedAcyclicGraphEdgesByBreadthFirstTree ( T ) for each directed edge e = ( q c , q a ) ∈ E do Obtain V a N = (cid:2) v ad − l +1 . . . v ad (cid:3) ∈ R d × l associated with the source subspace N q a M Obtain V c N = (cid:2) v cd − l +1 . . . v cd (cid:3) ∈ R d × l associated with the target subspace N q c M Define ¯ V a N = (cid:2) − v ad − l +1 v ad − l +2 . . . v ad (cid:3) ∈ R d × l Define ¯ V c N = (cid:2) − v cd − l +1 v cd − l +2 . . . v cd (cid:3) ∈ R d × l Define differentiable SO ( l ) R −→ a −→ cN , R −→ a ←− cN , R ←− a −→ cN , and R ←− a ←− cN , initialized near identity ( R −→ a −→ cN , L −→ a −→ c ) ← iterativelyMinimizeAlignmentError ( V a N R −→ a −→ cN , V c N ) ( R −→ a ←− cN , L −→ a ←− c ) ← iterativelyMinimizeAlignmentError ( V a N R −→ a ←− cN , ¯ V c N ) ( R ←− a −→ cN , L ←− a −→ c ) ← iterativelyMinimizeAlignmentError ( ¯ V a N R ←− a −→ cN , V c N ) ( R ←− a ←− cN , L ←− a ←− c ) ← iterativelyMinimizeAlignmentError ( ¯ V a N R ←− a ←− cN , ¯ V c N ) Associate ( R −→ a −→ cN , L −→ a −→ c ) , ( R −→ a ←− cN , L −→ a ←− c ) , ( R ←− a −→ cN , L ←− a −→ c ) , ( R ←− a ←− cN , L ←− a ←− c ) with e −→ r ) instead of flipped ( ←− r ): ori ( r ) = −→ r R rG = I l × l V Naligned ,r = V r N ori () of each point based on the minimum alignment loss, R G along the path to the point, V aligned N : Q = Queue () Q. enqueue ( childrenOfNodeInGraph ( r, E )) while size ( Q ) > do d = Q. dequeue () Q. enqueue ( childrenOfNodeInGraph ( d, E )) p = parentOfNodeInGraph ( d, E ) if L −→ d ori ( p ) < L ←− d ori ( p ) then ori ( d ) = −→ d R dG = R −→ d ori ( p ) N R pG V Naligned ,d = V d N R dG else ori ( d ) = ←− d R dG = R ←− d ori ( p ) N R pG V Naligned ,d = ¯ V d N R dG return { V aligned N associated with N q M for each q ∈ C M } .2 Additional ExperimentsA.2.1 Learning ECoMaNN on noisy data We also evaluate ECoMaNN learning on noisy data. We generate a noisy unit sphere dataset and anoisy 3D unit circle with additive Gaussian noise of zero mean and standard deviation 0.01. Afterwe train ECoMaNN on these noisy sphere and 3D circle datasets, we evaluate the model and obtain(82.00 ± ± P ¯ h M metric. These are still quitehigh success rates and outperform the VAE without noise.
A.2.2 Relationship between the number of augmentation levels and the projection successrate
We also perform an experiment to study the relationship between the number of augmentation levels(the maximum value of the positive integer i in the off-manifold points ˇ q = q + i(cid:15) u ) and the pro-jection success rate. As we vary this parameter at 1, 2, 3, and 7 on the Sphere dataset, the projectionsuccess rates are (5.00 ± ± ± ± The small positive scalar (cid:15) needs to be chosen sufficiently large as compared to the noise level, so that thedata augmentation will not create inconsistent data w.r.t. the noise.needs to be chosen sufficiently large as compared to the noise level, so that thedata augmentation will not create inconsistent data w.r.t. the noise.