Method of determining the safe navigation path of a autonomic robotic unit and robotic system navigating using this method

Abstract

Invention relates to a method of determining the safe navigation path of a autonomic robotic unit with only partial recognition of 3D object type and location during navigation of the autonomic robotic unit having a depth sensor when driving in a previously generated map of surrounding environment with unexpected appearance of unknown objects. Invention also relates to a robotic system comprising a autonomic robot comprising a depth sensor and central control unit navigating using a above-mentioned method.

Description

[0001] Method of determining the safe navigation path of a autonomic robotic unit and robotic system navigating using this method

[0002] Invention relates to method of determining the safe navigation path of a autonomic robotic unit and robotic system navigating using methods for navigation. More precisely invention relates to a unmanned autonomic robotic units navigating would human aid in constantly changing environment.

[0003] Background

[0004] For robots to navigate effectively, they must correctly estimate the passability of their surroundings. This is usually done by path planning in a 2D navigation plane, obtained by orthogonal projection of the obstacle space detected by the robot's sensors. This approach becomes a challenge when unexpected obstacles arise, such as sensors only seeing the surfaces closest to them. This makes it difficult to effectively estimate the size and shape of the obstacle, prohibiting immediate estimation of the occupied area.

[0005] Document EP3674983A1 discloses invention notably related to a computer-implemented method of machine-learning. The method comprises providing a dataset including 3D modeled objects which each represent a respective mechanical part. The dataset has one or more sub-datasets. Each subdataset forms at least a part of the dataset. The method further comprises, for each respective subdataset, determining a base template and learning a neural network configured for inference of deformations of the base template each into a respective 3D modeled object. The base template is a 3D modeled object which represents a centroid of the 3D modeled objects of the sub-dataset. The learning comprises a training based on the sub-dataset. This constitutes an improved method of machine-learning with a dataset including 3D modeled objects which each represent a respective mechanical part.

[0006] Document EP3674983A1 discloses a machine learning device includes: a generation unit generating a first shape model representing a shape of an object before deformation and a second shape model representing a shape of the object after the deformation based on measurement data before and after the deformation; and a learning unit learning a feature amount including a difference value between each micro region and another micro region that constitute the first shape model, and a relation providing a displacement from the each micro region of the first shape model to each corresponding micro region of the second shape model. Documents US20210125036A1 and JP2020077164 discloses a device, system for determining the orientation and shape of objects in an image using a neural network trained to determine the orientation of the object, and the use of an encoder-decoder system.

[0007] SUMMARY OF THE INVENTION

[0008] Invention relates to a method of determining the safe navigation path of a autonomic robotic unit with only partial recognition of 3D object type and location during navigation of the autonomic robotic unit having a depth sensor when driving in a previously generated map of surrounding environment with unexpected appearance of unknow objects events comprising steps: a. Acquiring data from a depth sensor of object as a partial cloud consisting of a set of points Xp generated at the intersection of the object's closest surface to the sensor and measurement beams originating from the sensor origin, bounded by the Field Of Vie, points Xp represents a surface of an object; b. Preprocessing of the cloud into a grid of voxels by wrapping the input point cloud into a cubic bounding box and tesselating it into a grid of n x n x n cubic voxels of a size r meters each; c. Calculating a cubic bounding box surrounding the grid of voxel having a center point; d. First 3D convolution of the bounding box using set of parameters: filter 16, kernel size 3x3x3 and stride 1 and max pool = 2; e. Preparing a first skip connection by 3D convolution of the first 3D convolution using filter 128 and kernel size 5x5x5; f. Second 3D convolution of the first 3D convolution results using set of parameters: filter 32, kernel size 3x3x3 and stride 1 and max pool = 2; g. Preparing a second skip connection by 3D convolution of the second 3D convolution using filter 32 and kernel size 5x5x5; h. Third 3D convolution of the second 3D convolution results using set of parameters: filter 64, kernel size 3x3x3 and stride 1 and max pool = 2; i. Preparing a third skip connection by 3D convolution of the second 3D convolution using filter 16 and kernel size 5x5x5; j. Resulting output from step h. is then compressed into fourth convolution 3D layer with 32 filters and then unrolled into a flat vector of size 16834 with dense layer forming a bottleneck comprising at least 8192 elements; k. Results of step j. are join with the first skip connection using cancat function; l. First upsizing is an operation operations of 16x16 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 16 filters; m. Results of step I. are join with the second skip connection using cancat function; n. Second upsizing is an operation operations of 32x32 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 32 filters; o. Results of step n. are join with the third skip connection using cancat function; p. Third upsizing is an operation operations of 64x64 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 64 filters;

[0009] Results of the step p. undergoes a final 2D convolution with a single filter which allows it to output a grayscale final image of a size n x n pixels corresponding to the footprint shape of a unknow objects the dimensions of the output image coincide with the size of a support square of the bounding box used for voxelization which is used for calculating the safe navigation path to avoid the unknow object.

[0010] Preferably the final image is thresholded with a parameter T which corresponds to a probability of a pixel being occupied.

[0011] Preferably depth sensor is a radio operated sensor - radar.

[0012] Preferably depth sensor is a light operated sensor - lidar.

[0013] Preferably depth sensor is a sound operated sensor - ultrasound sensor.

[0014] Invention also relates to a robotic system comprising a autonomic robot comprising a depth sensor and central control unit navigating using a above mentioned method.

[0015] The invention will include a method for estimating a 2D obstacle trace directly from incomplete 3D data in a point cloud file. Unlike existing point cloud completion methods, our approach does not require the reconstruction of the entire 3D object and its projection. Instead, it focuses on directly rendering a 2D representation relevant to navigation. The method is based on a lightweight, multimodal autoencoder that takes input from a voxelized incomplete point cloud and outputs a surface estimate that directly applies to the occupancy mesh. We evaluate our method on the PCN termination of the dataset and compare it with the results of point cloud completion algorithms projected onto a 2D plane. Our In most cases, this method proves to be superior in terms of accuracy and achieves the lowest resource consumption in all robotic pipeline testing.

[0016] Description of figures

[0017] The invention will now be described in more detail in a preferred embodiment with reference to the attached drawing, in which:

[0018] Fig. 1 Method uses an incomplete point cloud a), obtained by observing an object b) from a single perspective, to predict its footprint on the navigation plane d). This approach avoids the need to predict a complete cloud or drive around the object in order to estimate available space on the map. Without complete data, incorrect estimation of available space can occur c), resulting in inaccurate trajectory planning.

[0019] Fig. 2 System architecture of our proposal. An input partial cloud gets voxelized and fed into a 3D Convolutional Encoder. After the bottleneck, it is unrolled into a set of 2D Convolutions that directly predict a 2D footprint image on the navigation plane. Skip connections are used to provide a wider spatial context.

[0020] Fig. 3 Two ways to obtain complete footprint prediction. Top: current approach with intermediate point cloud completion, bottom: our approach with direct footprint estimation.

[0021] Fig. 4 Example of a preprocessed input cloud. A point cloud (left) undergoes binary voxelization to achieve discrete representation (right).

[0022] Fig. 5 Examples of results for various test samples and algorithms. 1st and 3rd row show point clouds, 2nd and 4th depict footprints obtained either by direct prediction (ours) or by projection (rest). One can notice how completion methods fail in case of low certainty resulting in multiple detached points which robots would treat as solid obstacles.

[0023] Fig. 6 Example of generation of a training sample. Ground truth clouds in the dataset are dense and fairly evenly distributed yet to ensure continuous footprint and no gaps within the contour (as it would be a solid, real-life object), we decided to use alpha shape approximation of cloud boundary with a high value a = 100 to include all details of the object's shape.

[0024] Fig. 7 Example of failure cases. Top row: input clouds, middle row: ground truth footprints, bottom row: output. Our method specifically fails in case of shape outliers (1st example) or tiny structures (3rd example). Embodiment

[0025] One of the key aspects of efficient robotic navigation is accurate perception, as it allows the robot to estimate safe trajectories that avoid obstacles. When driving in a previously generated map, the robot has complete knowledge of the traversability of its surroundings. However, if an unexpected object appears on the way, perception of its shape can be incomplete since the observation is done from a single viewpoint only. With sufficiently large obstacles, this prohibits accurate determination of the object's dimensions, effectively restraining the correct calculation of available space and possible trajectories. In such a case, only additional perspective would provide sufficient information, but it requires the robot to drive around the object, which is generally not desirable. In numerous applications, robots rely on 3D sensors that can only provide information about the surface closest to them, while other parts of the obstacle remain hidden. To estimate the complete shape of the object without additional motion, one needs to solve the problem of point cloud completion, also known as 'hallucination'. Many successful algorithms have been proposed to tackle this challenge, and their results could be used to predict missing parts of an object and augment the actual navigation pipeline. In practice, many robots convert perceived spatial information into a 2D occupancy grid by projecting 3D shapes onto a horizontal navigation plane. In such systems, performing whole point cloud completion is just a step in a pipeline that must be followed by a horizontal projection. This approach not only squanders spatial reconstruction precision (as the precisely reconstructed data is lost), but also leads to unnecessary expenditure of resources for additional operations. Instead of following such a path, this work proposes a solution that directly generates a predicted footprint from a partial 3D view of an obstacle (Fig. 1). The proposed system, coined PTFNet (Points To Footprint), revisits the concept of point cloud voxelization, which is suitable for different sensor types with various densities of point clouds. PTFNet is a multimodal autoencoder network that takes a voxelized partial point cloud as input and outputs the predicted area that the complete object would occupy directly on the navigation plane. The encoder part of the network uses 3D convolutions and skip connections, while the decoder part uses 2D convolutions and resizings to output a 2D estimation of the complete shape's footprint. Simple in structure, the network proves to be very accurate and allows for direct estimation of the occupancy grid without unnecessary computational overhead. This makes it specifically designated for robotic applications with limited onboard resources.

[0026] We evaluate our method on a well-known PCN completion dataset , originating from ShapeNet , for which we generate corresponding footprints of the complete object. We compare obtained results with 3D outputs of existing completion algorithms projected orthogonally on the navigation plane. Our method performs better than the majority of systems for various classes in terms of accuracy. It also outperforms all algorithms in terms of computational effort for the complete pipeline and comes second in the inference time.

[0027] To the best of our knowledge, our proposed system is the first of its kind to directly translate incomplete 3D data into a 2D contour representation. This innovation opens up numerous potential applications in the field of robotics, owing to its simplicity and efficiency. Our contribution to the field is twofold: Firstly, we present a novel, lightweight network that enables direct estimation of a 2D footprint from incomplete 3D point clouds. Secondly, we conduct a thorough numerical comparison of our approach with existing point cloud completion methods, thereby demonstrating the superiority of our proposed method.

[0028] Point Cloud Completion. Drawing navigation-relevant conclusions from incomplete 3D data is an underexplored field. So far, a full spectrum of methods focusing on incomplete 3D point cloud reconstruction has been proposed , but shape completion is often used without robotic context. The pioneering methods in this field, such as PointNet++ , TopNet , and PCN , propose direct operations on point clouds using Multi-Layer Perceptrons in various architectures. Multiple voxel-based methods have also been proposed, such as GRNet , which uses grids for the intermediate representation of clouds and others. These methods adopt a similar approach to ours by quantizing the 3D representation, but they propose much more complex architectures resulting in high computational effort. Recently, Transformer-based networks have been proposed in various variants, such as PointTr , an encoder-decoder architecture, or SnowflakeNet , which proposes progressive generation of points in the deconvolution step. The common characteristic of all these methods is that they aim for the most accurate completion of 3D without considering ways to utilize them in robotic applications, such as the need for projection and limited resources.

[0029] Occupancy Grid Estimation. Many works exist that draw conclusions about available space using RGB images. Watson et al. propose estimating it from a single image, but their focus is on predicting the entire area. In , a method for estimating support surfaces. However, these methods are not able to predict surfaces in the context of navigation occupancy grids. A similar approach to ours was proposed by Firman et al. , where they estimate occupancy from a single-shot depth image using Random Forest Prediction. However, their focus is on small objects and they aim for voxel (spatial) prediction. Choy et al. perform the opposite process, i.e., generation of voxel grids from single images. Recently, an interesting work was proposed by Kollmitz et al. , where the occupancy grid is predicted from collision events. Although this is not vision-related, it is an interesting example of applying neural networks for occupancy grid estimation.

[0030] Voxelised Representation. Using a voxelized approach for point clouds was proposed in in a form of efficient set of 3D Convolutions, applied for point cloud classification. Our encoder part uses similar approach but we differ in terms of the output representation. For multi-modal autoencoders some work explored them for pose estimation or for saliency prediction in depth and color data. None of these methods however employed 3D data to directly predict corresponding 2D images.

[0031] A depth sensor perceives an object as a partial cloud consisting of a set of points Xpgenerated at the intersection of the object's closest surface to the sensor and measurement beams originating from the sensor origin, bounded by the Field Of View of the sensor. A robot navigating by means of 2D occupancy grid would project these points orthogonally on a horizontal plane.

[0032] Assuming that the origin of the partial cloud is known and coincides with the centroid of the complete object (which is the assumption every point cloud completion algorithm adopts), one can define orthogonal projection in the following way. A footprint can be expressed as a binary image Y of dimensions n x n pixels, composed of squared pixels of size corresponding to the resolution of the navigation map i.e. r meters. A partial point cloud is projected such that the origin of it lies in the centre of the projection image. More specifically, let p'~ be an ith point of Xpand Yp(u,v) be a 2D footprint image with pixel coordinates u,v E Z+and origin in top, left corner. For each point p' the coordinates of the corresponding projection (occupied) pixel can be calculated as follows:

[0033] The footprint image is defined as a binary projection of all points in the point cloud Xpon the 2D horizontal plane such that Ypuocc,vOcc) = 1. The remaining, non-occupied pixels are set to YP(unocc,Vnocc) = 0.

[0034] If the entire object's shape is known one could obtain the complete footprint ycby a similar projection. In this way the task can be realised by completing the point cloud using one of existing algorithms and projecting the result on the navigation plane. In this paper we instead claim that it is possible to directly estimate footprint of complete object ycfrom partial point cloud Xpwithout intermediate point cloud completion step i.e. without knowledge of the whole 3D shape. The comparison of two approaches is depicted in Fig. 3. Our method assumes that the input point cloud represents a surface of an object therefore in a robotic pipeline it should be preceded by a segmentation mechanism e.g. or.

[0035] Our algorithm requires a preprocessing step of the input cloud and has a structure of a multi-modal autoencoder. The overview is depicted in Fig. 2 and the details are described in following sections.

[0036] Preprocessing

[0037] The system requires discrete representation in a form of binary voxel grid. The grid is created by wrapping the input point cloud into a cubic bounding box and tesselating it into a grid of n x n x n cubic voxels of a size r meters each. A cloud is placed within the bounding box so that its origin (which coincides with the centroid of a complete object as stated in section III) is located in the centre of the voxel grid. A voxel is considered occupied if there exists at least one point within this voxel. An example of preprocessed input is depicted in Fig. 4.

[0038] Network

[0039] The detailed architecture of our proposed network is depicted in Fig. 2. As shown there, it has a structure of a convolutional encoder-decoder augmented with additional skip connections. A fundamental feature of our approach is the multi-modal data representation it uses. The input format, on which the encoder operates, is a discrete representation of 3D spatial distribution of points (voxel grids). Consequently, the decoder operates directly on 2D data (images). Such an approach allows us to grasp the influence of various elements of a spatial shape on resultant projection in the absence of complete structure. The alternative approach would be to directly project the (partial) point cloud and operate only on 2D images to achieve missing footprint shape estimation. This approach would however inevitably lead to the loss of significant spatial information.

[0040] The encoder part of the network consists of three stages of 3D convolutions of 16, 32 and 64 filters respectively with kernels of size 3 x 3 x 3 and strides = 1 each. Each stage is concluded with Max pooling of pool = 2 Output of the encoder is compressed into another 3D convolution layer with 32 filters and then unrolled into a flat vector of size 16834 densely connected to a bottleneck of 8192 elements. Next, the decoder follows with three stages of upsizing operations of 16x16, 32x32 and 64x64 each using nearest neighbour interpolation method. Each upsizing step is followed by 2D convolutions with kernels 3 x 3 and strides = 1 of 64, 32 and 16 filters respectively. Parameters of these convolutions (the main branch of the network) are selected specifically to catch tiny details of spatial distribution of an object's shape and its influence on the resultant projection.

[0041] On the other hand, to achieve more robust estimation capabilities of the network, a broader context of the shape is required as the main branch misses the correlation between more distant pieces of the shape. Consequently the network is expanded with additional skip connections. These skip connections are implemented as 3D convolutions layers with bigger kernel sizes (5 x 5 x 5) with 128, 32 and 16 filters respectively. The first skip branch is connected directly after the bottleneck. Remaining two concatenate the output of a respective 3D convolution with the output of each resizing operation, right before 2D convolution step.

[0042] The output of the decoder undergoes a final 2D convolution with a single filter which allows it to output a grayscale image of a size n x n pixels corresponding to the footprint shape. The dimensions of the output image coincide with the size of a support square of the bounding box used for voxelization. The final image is thresholded with a parameter T which corresponds to a probability of a pixel being occupied. For all convolutions we used same padding. Our model has total 135,947,953 parameters with 17,248 being non-trainable.

[0043] Loss Function

[0044] The network is trained by penalizing the difference between the Ground Truth projection of a complete shape / crand the obtained prediction Yp, which implies that loss is computed solely in the 2D domain (fixed-size images comparison). Additionally, the images are binary, with pixels corresponding to free or occupied space. We propose to verify two candidates for the loss function. The first one is a classical Mean Squared Error metric: where i and j correspond to pixel coordinates. Since the MSE takes into account the complete image, in case of small objects (footprint occupying minor part of image) the penalty for mismatch might be less strict as the background reduces overall loss score. To mitigate that we propose another loss function namely Jaccard Loss (Intersection Over Union) expressed in the continuous domain

[0045] With e preventing division by 0. This loss function is broadly used in segmentation tasks and in that regard is relevant to our application as it considers only the actual contour of the footprint. In our experiments we validate both functions.

[0046] EXPERIMENTS To validate our method we performed multiple experiments described in this section.

[0047] Baselines

[0048] In the absence of methods predicting footprints directly from partial point clouds we contrast our method with previously proposed algorithms according to the rationale depicted in Fig. 3. Solutions used in the comparison include several newly proposed methods namely SnowFlakeNet , PointTr , SpareNet , PMP-Net++ , GRNet as well as the pioneering Point Completion Network.

[0049] Fig. 6: Example of generation of a training sample. Ground truth clouds in the dataset are dense and fairly evenly distributed yet to ensure continuous footprint and no gaps within the contour (as it would be a solid, real-life object), we decided to use alpha shape approximation of cloud boundary with a high value a = 100 to include all details of the object's shape.

[0050] To generate training datasets, footprints were obtained by projecting complete point clouds on the horizontal plane, applying alpha shape approach for obtaining continuous contour and discretizing resulting images into pixels (Fig.6). For numerical evaluation we use the Intersection Over Union metric, which specifically catches how well the shape of the footprint was reconstructed without considering the background. Since, both the projections and the outputs of our network are binary images, the loU metrics takes a simple form as follows:

[0051] For our experiments we use the images of a size n x n = 64x64 pixels. All data samples that we used are normalized (fitting in a unit bounding box) however the choice of image size was made to comply with robotic applications. With resolution of a map r = 10cm an image would correspond to 6.4m which can reasonably catch real-life object dimensions. Same size (n = 64) was selected for the voxel bounding box.

[0052] PCN dataset

[0053] The dataset used for validation was proposed by the authors of Point Completion Network. It originates from Shapnet and consists of 30974 complete point clouds of 8 different classes of object i.e. Plane, Cabinet, Car, Chair, Lamp, Sofa, Table, Vessel. For each complete point cloud, 8 partial clouds are generated from various viewpoints. We follow the train, validation and test split originally proposed by authors of PNC. For the purpose of our validation we generated footprints for complete point clouds and voxelized all partial point clouds in the dataset. Consequently, we run multiple experiments to validate the accuracy of our system and to compare the results of projection of other algorithms using the metric from Eq. 4, the results are summarised in Table I. Additionally, due to the target of applicability of our solution we also evaluated the computational effort of various algorithms, the results are gathered in Table II. Finally the exemplary visual outputs are depicted in Fig.5.

[0054] Fig. 7: Example of failure cases. Top row: input clouds, middle row: ground truth footprints, bottom row: output. Our method specifically fails in case of shape outliers (1st example) or tiny structures (3rd example). It also struggles with disconnected shapes.

[0055] Our method outperforms other algorithms in the majority of cases by a noticeable margin. The best performance can be noticed for solid (typically convex) structures e.g. Cor or Table. The failures happen specifically for tiny structures e.g. Lamp or in case of separated structures in one test cloud (for details of failure cases refer to Fig. 7). Our experiments also lead to interesting observations that results of cloud completion networks often are noisy which get propagated to corresponding footprints (see middle rows in Fig.5). One way to mitigate this issue would be to use similar approach as we did with the training dataset generation namely to apply alpha shape approach on the projected point cloud but this method fails when a point cloud has noticeable variations of density and will effectively filter out all separated points (e.g. the detached points in last row of 5), therefore we decided not to follow that approach.

[0056]

[0057] TABLE I: Results of prediction of our algorithm compared to the results of projection of other methods, comparison metric is loll. PTFNet outperforms other methods in the majority of classes and comes ex aequo with Snowflake in terms of average performance. It can be also noticed that Jaccard loss provides slightly better results than MSE but the difference is not very significant. only two steps: pre-processing and prediction and in total (average) takes less than any other method (~11 ms vs ~43 ms of the fastest PCN). The other methods require pre- and post-processing. We accumulated the times of those in single columns as these operations for every method are very similar. One should notice that pre- and post-processing depends heavily on the number

[0058] Furthermore, we conducted detailed experiments to measure the time performance of all methods.

[0059] Table II summarizes the results, including inference times and pre- and postprocessing times. Our approach requires only voxelization of the input cloud and no post-processing. Other algorithms, on the other hand, require adjusting the cloud size to k points at the beginning and projection after inference. To measure preprocessing time we used implementations proposed by other methods, for postprocessing we implemented it ourselves in a similar way as the voxelization. Our method outperforms all others when considering the entire pipeline, and ranks second in inference times, behind PCN, which achieves those results at a cost of lower accuracy. Saving resources is critical in robotics, making our approach a promising solution.

[0060] Training setup

[0061] We conducted all experiments on a machine with Zotac RTX 3090 (24GB RAM - 10496 CUDA cores) GPU (learning and prediction) and with AMD Ryzen 7 3700x(8 cores) CPU (pre- and post- processing). Our entire pipeline was implemented using Tensorflow 2.6.2. We setup the training for 25 epochs with early stopping callback of patience 4, data was fed in batches of 8. Each layer in the network was of points in the cloud activated using Leaky ReLU with a = 0.1 and concluded by Batch Normalization with zero momentum.

[0062] We identify three main components on which the performance of our algorithm depends, we compared the results using loU metric gathered them in the Table III.

[0063] 3D Convolutions. To judge the influence of spatial distribution of points on the footprint we modify the encoder so that the input point cloud gets projected before entering the network and all 3D Convolutions are changed to 2D with corresponding numbers of filters and kernel sizes. As the results suggest, the loss of spatial information upfront deteriorates the performance significantly; the projection is not informative enough to catch missing parts of an object.

[0064] Skip Connections. To evaluate the influence of skip connections we tested a simplified version of our algorithm i.e. "vanilla autoencoder" which consists of only the main, horizontal branch from Fig. 2. The results are poorer but still satisfactory, suggesting that spatial context resulting from bigger kernels of skip connections improves the general prediction capabilities.

[0065] Bottleneck size. The last parameter that could potentially influence the performance is the size of the bottleneck. As the table shows reducing the bottleneck size noticeably reduces the number of parameters without significant drop in performance. This could be another way to further optimize computing resources.

[0066] TABLE III: Results of ablation study

[0067] We showed that it is possible to accurately estimate an obstacle footprint from various samples of incomplete data and do so without unnecessary resources expenditure. Our method proves successful on various objects and offers an output relevant for robotic context in a time-efficient manner. The system also generalizes well on different classes. The results indicate a high potential of our method for robotic applications however it requires further verification on real life data.

Claims

Claims1. Method of determining the safe navigation path of a autonomic robotic unit with only partial recognition of 3D object type and location during navigation of the autonomic robotic unit having a depth sensor when driving in a previously generated map of surrounding environment with unexpected appearance of unknow objects events comprising steps: a. Acquiring data from a depth sensor of object as a partial cloud consisting of a set of points Xp generated at the intersection of the object's closest surface to the sensor and measurement beams originating from the sensor origin, bounded by the Field Of Vie, points Xp represents a surface of an object; b. Preprocessing of the cloud into a grid of voxels by wrapping the input point cloud into a cubic bounding box and tesselating it into a grid of n x n x n cubic voxels of a size r meters each; c. Calculating a cubic bounding box surrounding the grid of voxel having a center point; d. First 3D convolution of the bounding box using set of parameters: filter 16, kernel size 3x3x3 and stride 1 and max pool = 2; e. Preparing a first skip connection by 3D convolution of the first 3D convolution using filter 128 and kernel size 5x5x5; f. Second 3D convolution of the first 3D convolution results using set of parameters: filter 32, kernel size 3x3x3 and stride 1 and max pool = 2; g. Preparing a second skip connection by 3D convolution of the second 3D convolution using filter 32 and kernel size 5x5x5; h. Third 3D convolution of the second 3D convolution results using set of parameters: filter 64, kernel size 3x3x3 and stride 1 and max pool = 2; i. Preparing a third skip connection by 3D convolution of the second 3D convolution using filter 16 and kernel size 5x5x5; j. Resulting output from step h. is then compressed into fourth convolution 3D layer with 32 filters and then unrolled into a flat vector of size 16834 with dense layer forming a bottleneck comprising at least 8192 elements; k. Results of step j. are join with the first skip connection using cancat function; l. First upsizing is an operation operations of 16x16 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 16 filters; m. Results of step I. are join with the second skip connection using concat function;n. Second upsizing is an operation operations of 32x32 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 32 filters; o. Results of step n. are join with the third skip connection using cancat function; p. Third upsizing is an operation operations of 64x64 using nearest neighbour interpolation method followed by 2D convolutions with kernels 3 x 3 and strides= 1 of and 64 filters;Results of the step p. undergoes a final 2D convolution with a single filter which allows it to output a grayscale final image of a size n x n pixels corresponding to the footprint shape of a unknow objects the dimensions of the output image coincide with the size of a support square of the bounding box used for voxelization which is used for calculating the safe navigation path to avoid the unknow object.

2. Method according to claim 1, characterized in that the final image is thresholded with a parameter T which corresponds to a probability of a pixel being occupied.

3. Method according to claim 1 or 1, characterized in that depth sensor is a radio operated sensor - radar.

4. Method according to claim 1,2 or 3, characterized in that depth sensor is a light operated sensor - lidar.

5. Method according to claim 1,2,3 or 4, characterized in that depth sensor is a sound operated sensor - ultrasound sensor.

6. A robotic system comprising a autonomic robot comprising a depth sensor and central control unit navigating using a method according to any of claims form 1 do 5.

Images