Transfer Learning of Traversability Assessment for Heterogeneous Robots

Abstract

For autonomous robots operating in an unknown environment, it is important to assess the traversability of the surrounding terrain to improve path planning and decision-making on where to navigate next in a cost-eficient way. Specifically, in mobile robot exploration, terrains and their traversability are unknown prior to the deployment. The robot needs to use its limited resources to learn its terrain traversability model on the go; however, reusing a provided model is still a desirable option. In a team of heterogeneous robots, the models assessing traversability cannot be reused directly since robots might possess diferent morphology or sensory equipment and thus experience the terrain diferently. In this paper, we propose a transfer learning approach for convolutional neural networks assessing the traversability between heterogeneous robots, where the transferred network is retrained using data available for the target robot to accommodate itself to the robot's traversability. The proposed method is verified in real-world experiments, where the proposed approach provides faster learning convergence and better traversal cost predictions than the baseline. heterogeneous robots, transfer learning, neural networks, traversability assessment CEUR

1. Introduction

Our work is motivated by autonomous tasks such as mobile robot exploration, where robots encounter terrains that might impede their movement but have unknown properties due to the nature of the mission. In such deployments, the robots can improve the eficiency of their navigation by learning the terrain properties incrementally during the mission. Further, we can reason about distributing the exploration to multiple robots to finish the mission faster. For a heterogeneous team of robots, each robot can be assigned to a suitable part of the environment, such as a small crawler exploring tight spaces, while larger and faster robots can be assigned to open areas. However, robotic platforms with varying builds and sensory equipment have diferent terrain perceptions; hence, each platform needs to learn standalone terrain assessment models. The knowledge transfer approach can reduce the complexity of training and maintaining multiple standalone models.

Transfer learning is a part of the machine learning principles that aim to improve the performance in the tarCEUR htp:/ceur-ws.org ISN1613-073 1The source domain (teacher) denotes the entity providing knowlTeacher's model

Transferred student's model Terrain observation

Transfer of model + Continuous learning

Terrain observation using different

sensors Predicted cost

Predicted cost teacher is available, the teacher’s model is transferred to the student and modified to accept the student’s observation format indicated by the diferent colormaps. Then, the transferred model continues learning using the student’s observations to be informed about the student’s experiences.

In this paper, we propose to utilize transfer learning to share terrain traversability assessment models between heterogeneous robotic platforms, as illustrated in Figure 1. The individual models are neural networks that predict a continuous score describing the dificulty of the ized by weights transferred from the teacher, and then the student’s target domain. If the dimensions of observations are unequal due to diferent sensors being used by the teacher and student, the input dimensionality is reduced (or increased) by additional neural network layers. The proposed approach is compared to the baseline, learned only in the target domain, using a robot with edge to the individual in the target domain (student). get domain by the experience in the source domain [ 1 ]1. terrain traversal. The student’s neural network is initialITAT’22: Information technologies – Applications and Theory, Septem- the neural network is tuned using the data obtained in © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License heterogeneous terrain experience simulated by various Attribution 4.0 International (CC BY 4.0). traversability assessment methods. Besides, the feasi- be employed as in [ 13 ], where it is advocated that the bility of the approach is validated in the experimental learning rate of the neural network is set lower when scenario with real heterogeneous robots. applying only slight corrections to weights during the

The rest of the paper is organized as follows. The fine-tunning. While fine-tuning, it can be beneficial to rerelated work on traversability assessment and transfer frain from updating the weights in some layers (denoted learning is briefly reviewed in Section 2. In Section 3, the as layers freezing) [ 14 ]. problem of traversability transfer between heterogenous The costly data collection and various tasks and robots’ robots is presented. Section 4 describes the proposed bodies make robotics an interesting field for deploying approach for the transfer of traversability assessment transfer learning techniques. Although testing only in experience. The evaluation results of the proposed ap- simulators, the authors of [ 15 ] utilize transfer learning proach are presented in Section 5. Section 6 concludes to propagate experience in diferent scenarios of robotic the paper. soccer, where they propose a solution for transferring neural networks between tasks with diferent inputs and action spaces. In [ 16 ], humanoid robots observe human 2. Related work gestures and motions to replicate them later. Another method to transfer human experience to neural networks utilized by robots is presented in [ 17 ], where humans provide knowledge directly to the network. A robotic arm is trained to reach a destination of a colored block in [ 18 ]. The transfer is carried out between robotic arms with a diferent number of joints.

However, since the traversability over a single terrain may greatly vary between robotic platforms, such as wheeled and legged ground vehicles [ 19 ], the obtained knowledge cannot be shared directly between diferent robots. Therefore, we aim to utilize transfer learning to distribute a traversability assessment model consisting of a neural network.

The traversability assessment is to support path planning and decisions, such as avoiding impassable terrain or optimizing the path for the specific needs of the robot [ 2 ].

Correct traversability estimation is essential for applications where the robot encounters diferent terrains, including dangerous environments. Such fields are represented by extra-terrestrial exploration [ 3, 4, 5 ], search and rescue missions [ 6 ], and agriculture or of-road driving [ 7 ]. In [ 2 ] and [ 8 ], the authors provide a thorough overview of traversability assessment methods suitable for mobile robots. Therefore, we focus our brief review on recent neural network approaches, which provide appearance-based traversability predictions and utilize image-processing and classification methods.

In [ 5 ], a fully convolutional neural network is utilized 3. Problem Statement to locate the best possible place for the rover to land by classifying multiple terrain types and has been used We examine various robots perceiving diverse terrains for the Mars rover mission. An alternative architecture during operational usage. Let the robots be deployed is proposed in [ 9 ], where traversability predictions on in an environment modeled as the 2.5D grid, where each future paths are achieved using Generative Adversarial cell can be labeled by a number, and thus ∈ ℕ , and the Networks (GANs) [ 10 ] that create virtual images from cell size corresponds to the footprint of the smallest the already traversed path. However, a vast dataset is robot in the team. The center of each robot’s footprint is needed to train a neural network from scratch, fully high- discretized as the cell robot ∈ ℕ. Robots move through lighting the need for an approach capable of enhancing the environment along paths that are represented as performance with a smaller dataset. sequences of neighboring cells 1, … , corresponding to

For such cases, transfer learning can be employed, the robot’s discretized positions. which is an approach capable of improving knowledge The robot ’s path-planning decisions are made with in the target domain by the transfer from the source do- respect to (w.r.t.) the particular robot’s cost by finding main [ 1 ]. In [ 11 ], the authors utilize transfer learning a path with the minimal expected cost in the form of the weight transfer from remotely similar tasks to reduce the dataset size in the target domain ∗, = argmin ∈Ψ(, ′) ∑ ( ), (1) necessary to train the convolutional neural network and ∈ thus shorten the training time. After transferring weights where Ψ(, ′) is a set of all paths from to ′. Howbetween tasks, it is desirable to fine-tune them to suit ever, the cost function is not known a priori; thus, the the task’s needs in the target domain. The classification robot has to learn it to estimate the cost by ̂ . Hence, feature extractor output layers are reinitialized in [ 12 ] a traversability assessment model is needed that asto comply with the possibly diferent classes in the tar- signs the predicted cost ̂ for a terrain observed using get domain. If the source and target tasks vary greatly, exteroceptive sensors as an entire redesign of the classifier’s architecture may ∶ → ̂ , (2) where is the observed terrain appearance.

which each robot can compute using its cost computaSince the mobile robot’s traversability is considered tion method. All the cost computation methods return too complex to be assessed using a handcrafted function, a model is trained from the robot’s experience to predict its future cost ̂ . The costs utilized for training are computed using proprioceptive sensors because of their ability to measure how the environment influences the robot’s body. The training of each traversability model aims to minimize the Root Mean Square Error (RMSE) between the model’s cost assessments and the cost measured using proprioceptive sensors as √ 1 =1 RMSE = ∑( (( )) − )2.

(3)

We address the traversability assessment for heterogeneous robotic platforms 1 ≠ 2, which possess diferent capabilities. Therefore, we assume that diferences between the platforms can result in unequal cost measurements

1 ≠ 2 over some terrain . On the other hand, we assume that there are underlying similarities between how the robots interact with the terrain. The problem being addressed is to improve the performance of traversability assessment by transferring the cost assessment model

1 from the robot 1 before learning on the robot 2, and thus learning its cost assessments relatively sooner while achieving similar or better predictions than the regressor 2 trained using only the 2’s data. 4.

Method

In the proposed approach for transferring mobile robot terrain traversal experience between the heterogeneous robots, the traversability experience is denoted as the traversal cost . Then, for each robot, the costs are predicted using the robot’s regressor . Each regressor is a neural network trained using the robot’s prior traversal costs associated with the description of the particular terrain where the cost was experienced. The teacher’s terrain experience, represented as the teacher’s learned network, is transferred to the student who has no prior .

cmd strictly positive values because zero and negative values would incentivize infinite paths, preventing the robot from reaching its goal. We consider the following cost computation methods.

Velocity - monitors relation of the achieved speed and commanded velocity using equation = Slope - computes the cost = 1 + as an angular distance in the degrees from the straight pose . The ofset by 1 is to accent the energy expenditure even on flat terrains.

Diference

- is defined similarly to as = 1 + , where expresses the maximal angular distance in degrees of the subsequent robot’s positions. Further, the costs are adjusted as adjusted = max ⋅ tanh (4) where stands for , , and . The adjustment is to lower the high-cost values for cases the achieved velocity is negligible when compared with commanded velocity because the robot gets stuck.

Tilt - is computed as = = | tan |, where is the absolute angle of the two opposing footholds (for the legged robots) from the flat surface. For example, the left front and right rear legs are considered opposing. The values and a diference in elevation and flat-plane distance measure of the footholds, respectively.

4.2. Traversal Cost Regressor

The cost regressor , based on the regressor proposed in [ 20 ], uses the terrain appearance and geometry to assess the robot’s traversal cost. The regressor is a neural network that uses the image processing-like architecture

max terrain experience by using the teacher’s weights to ini- shown in Figure 2 and operates as follows. tialize the student’s network. After the transfer, the student’s network is further trained to adapt to the student’s domain fully.

In the rest of this section, we describe in detail the

During the deployment, the robot uses its range measurements to build a height map ℕ of the mission environment in the form of an elevation grid map with the squared cell size of 7.5 cm [ 21 ]. Depending on the carried traversal costs used by the individual robots, the regres- sensory equipment, the grid map may also include the tersor, its learning process, and the knowledge transfer.

4.1. Robots’ Traversal Costs

The traversability assessment regressor is trained on trails that include observations of the traversed terlocalized in ℕ paired with the terrain observations at the respective locations. Each terrain observation is in the rain paired with the cost perceived over the terrain, form of a × × segment centered at the location, where rain color in addition to the elevation information, which can be further utilized in regression and extrapolation of the learned traversal experience.

The regressor is learned from the cost measurements Terrain observation conv1

conv2 8 x 8 x n

5. Results

The proposed knowledge transfer method has been examined in several experimental scenarios. First, we simulate the heterogeneity of the robots using a small hexapod crawler with varying cost perception, which provides an easy way to verify the feasibility of the proposed approach. Then, we display the proposed knowledge transfer using two diferent real robots.

(a) (b) is the observation window size selected so that the entire robot’s footprint is covered. The dimensionality is either 1 when only range measurements are available, or 3 in the case range measurements are accompanied by a and b channels of the lab color space. The network is learned using Adam optimizer w.r.t. the mean absolute percentage error

| Λ = 100 ||( − ) 1 |

| , | (5) where is the expected output of the neural network and the prediction.

4.3. Knowledge transfer

During the knowledge transfer, the weights from the The proposed method can be used with any set of source domain are utilized in the target domain. Besides, ground vehicles. However, we focus on multi-legged when needed, the transferred model is adjusted as shown robots since their traversal capabilities permit deployin Figure 3. The width of the input window is prepared ment in a wide range of terrains. The utilized robots separately since it is selected to fit the robot receiving are the small hexapod crawling robot SCARAB II [ 22 ] the model. However, the observation dimensionality can and the four-legged Spot that are depicted in Figure 4. difer between teachers’ and students’ exteroceptive sen- Besides their morphology, the robots also difer in size. sors. In such a case, an additional convolutional layer is Spot’s footprint is larger than SCARAB II that occupies used to reshape the input and accommodate the trans- a disk with a 25 cm radii, while Spot’s footprint fits into ferred regressor to the student’s perceived data. The layer 1.1 m × 0.5 m rectangle. Furthermore, SCARAB II carried comprises 1 × 1 convolutional kernel with the input and the Intel RealSense Tracking Camera T265 and the RGB-D output channels corresponding to the perceived number Intel RealSense Depth Camera D435 providing depth and of channels and the number of transferred regressor’s color appearance exteroceptive data. Spot perceived only input channels, respectively. range measurements using the Ouster OS0-128 LiDAR

The new model is retrained using the student’s dataset and does not perceive color. collected by the student. It is because the teacher’s and student’s costs are heterogeneous, albeit it is assumed 5.1. Cost Assessment Methods that the network captures underlying terrain properties Examination that can be transferred. Additionally, during the retraining of the regressor, layers can be frozen by fixing their The feasibility of the proposed method is firstly veriweights since it is assumed that the initial layers extract ifed in a scenario where the diference in perception of general features that are primarily similar between vari- heterogeneous robots is simulated using various cost asous data. sessment methods of SCARAB II. The robot collected the datasets in the Bull Rock Cave near Brno, Czech Repub- Table 2 lic, as described in [ 20 ]. The datasets were collected in various parts of the cave system, and each set is a result of the robot walking over one of the particular terrains, whose selection is shown in Figure 5. Each dataset collection included approximately 5 minutes of the navigation, enabling the robot to observe 6 × 6 m of the environment.

RMSE’s mean (std) values between 300 and 600 epochs of → transfer.

Epochs

300 600

In Table 2, the transfer from to is chosen to ex- training for 300 epochs, the RMSE of the regressors’ preamine the influence of training for increased number of the training epochs. The same randomly generated transfer scenarios are utilized as in Table 1; however, the dictions against the collected ground truth is 3.69 and 1.84 for the direct and transferred models, respectively.

Hence, we can conclude that the transferred model immodels are trained for 600 epochs instead of 300. The re- proved performance faster. sults show that the direct student’s model has improved with the increased number of epochs, while the RMSE of the transferred model slightly increased during the prolonged training, likely because of overfitting the training data. The transfer model learned with 300 epochs has achieved suficient performance, and the results suggest that the transfer helps reduce the necessary number of

Cost Assessment Method 5.1.1. Transfer between Slope and Velocity - We further examine the transfer between the student’s slope and teacher’s velocity cost assessment method in detail as those methods compute cost using dissimilar approaches. The dataset is split so that the student’s dataset is overall a third size of the teacher’s, hence suitable to showcase the knowledge transfer as there is much information to be received by the student. Teach90 80 sso L70 60 50

Steps 30

Direct Transfer + Direct Ground truth 40

50 with artificial grass and spikes in the form of soundproofing material. Outdoors, the robot traversed various surfaces such as hard-packed soil, cobbles, and sloped grass.

Spot can move faster than SCARAB II; thus, the Spot’s datasets are longer, as Spot traverses more terrain in similar 5-minute long deployment, where Spot is capable of traveling through 15 × 15 m environment. Therefore, using fewer datasets to train the cost assessment model is suficient. In the following paragraphs, we examine the performance when transferring both from Spot to SCARAB II, and in the opposite direction.

5.2.1. Spot Knowledge Transfer to SCARAB II

Moreover, we examined the traversal of a single cave - The transfer from Spot to SCARAB II is achieved ustrail. Figure 7 shows the predicted costs by the student’s ing observation windows with the width of = 8 cells, direct and transferred models compared to the collected which is suitable for the smaller hexapod crawler. Each ground truth. We can observe that the transferred model transfer scenario, comprising transfer from Spot to one of follows the ground truth better than the direct model. the hexapod’s cost models, is evaluated in 5 setups. For each setup, 5 datasets are randomly chosen to train the

Spot’s teacher model, while SCARAB II receives 6 ran0 Co5st 10 0 Co5st 10 0:25Heigh0t:0[m0] 0:25 domly chosen datasets. The trained models are examined 0 0 0 on 5 randomly chosen datasets, which difer from the training sets. The regressors are trained for 300 epochs lleC2400 lleC2400 lleC2400 witShCaAR9-AtoB-1II tmraoidneinlsgt-hveaelindvaitrioonnmspelnitt.as a colored height 60 60 60 map, while Spot uses only a height map. Thus, we con0 20 Cell 40 0 20 Cell 40 0 20 Cell 40 sniedlesritnheeastcuhdternatn’ssfmerodsecleinnapruiot.wWithhbeonthu s=in g{1t,h3e} tchhraene-(a) Direct model (b) Transfer model (c) Height Map channel version, which perceives both the elevation and Figure 8: (a) Direct and (b) transferred model’s cost assess- the a and b channels of the lab color space, a convoluments of the perceived environment after training for 300 tional layer reshaping the input, is added to accommodate epochs; (c) and the environment’s height map, where the path teacher’s (Spot’s) model that has only one input channel. of the robot is in red. The shown maps have squared cells with the size 7.5 cm.

5.2. Transfer between SCARAB II and

Spot

The dataset comprising heterogeneous robots is created by adding the Spot’s datasets to the SCARAB II’s data. The datasets were collected in indoor and outdoor locations of the Czech Technical University in Prague campus at Charles Square, Prague, Czech Republic. Indoors, see Figure 4b, Spot moved over surfaces partially covered

Table 3 shows the performance of the trained models. All transferred models perform overall better than the direct model. However, the performance of the transferred model has not improved when modifying the teacher’s model to accept the colored height map collected by SCARAB II, although the direct model has improved when using = 3 input channels. In the authors’ opinion, the added convolutional layer could not suficiently modify the input observation to achieve good performance in combination with the underlying transferred model.

5.2.2. SCARAB II Knowledge Transfer to Spot

- For the transfer from SCARAB II to Spot, the scenarios are adjusted by using = 16

to match Spot’s body size, and the regressors are trained for 100 epochs in tional layer is added during the transfer to Spot to utilize Spot’s target domain. Besides, the reshaping convolu- for 300 epochs. the SCARAB II’s model with the three input channels. 6 tso 4 C 2 0 0 200 40E0pochs the direct model. However, even the transferred model cannot closely follow the oscillations of the ground truth. (a) Direct model (b) Transfer model channels .

Observ.

Depth Depth+ ab Scenario

5.3. Individual Transfer between Spot and SCARAB II

We further present a detailed overview of the knowledge transfer between Spot and SCARAB II. SCARAB II utilizes the diference cost computation method , and the respectively.

The progress of the regressors’ training after a prolonged training for 1000 epochs is depicted in Figure 9a. width of the observation window is set to = 8 . Af- sessments of the perceived environment after training for ter 300 training epochs, the student’s direct transferred 300 epochs; (c) and the environment’s height map with the and fine-tuned models achieved RMSE of 1.79 and 1.09, squared cell of the size 7.5 cm.

Figure 10 illustrates the cost assessments and the 0 50

Cost 5

10 100 Cell

150 0 25 ll e50 C 75 0 0 50 0 25 ll e50 C 75 0 Heig1ht [m] 2 100 Cell

150 (c) Height map 0 50

Cost 5

10 100 Cell 150 height map with the marked robot trail in the Room part of the cave. Since the robot traversed only a tiny part of the observed environment, we present a manual evaluation of the cost assessments as the cost measurements are not presented in the observed area. The transferred model assigns a higher cost to the terrain edge, while the student’s model underestimates the dificulty. Additionally, the transferred model suggests a more challenging cost in all areas of the environment, which resembles the actual perceived cost more closely. Thus, we conclude that the transfer can correct the student’s direct model predictions.

6. Conclusion

In this paper, we present an approach for sharing knowledge about traversability between heterogeneous robots. Traversal cost predictors are created using neural networks processing observations from exteroceptive sensors. The knowledge transfer is implemented as the transfer of neural network weights, and the transferred networks are fine-tuned to adapt to the receiving robot’s terrain perception. The proposed method is verified using a small hexapod crawler and a large quadruped walker, with the proposed method lowering the traversability prediction error. Next, we aim to deploy the proposed method in path planning tasks, with the final goal of simultaneous online learning on multi-robots.

Acknowledgments

This work was supported by the Czech Science Foundation (GAČR) under research project No. 21-33041J.

References

[1] S. J. Pan , Q. Yang , A survey on transfer learning , IEEE Transactions on Knowledge and Data Engineering 22 ( 2010 ) 1345 - 1359 . doi:1 0 . 1 1 0 9 / T K D E . 2 0 0 9 . 1 9 1 .

[2] P. Papadakis , Terrain traversability analysis methods for unmanned ground vehicles: A survey , Engineering Applications of Artificial Intelligence 26 ( 2013 ) 1373 - 1385 . doi:1 0 . 1 0 1 6 / j . e n g a p p a i . 2 0 1 3 . 0 1 . 0 0 6 .

[3] S. Singh , R. Simmons , T. Smith , A. Stentz , V. Verma , A. Yahja , K. Schwehr , Recent progress in local and global traversability for planetary rovers , in: IEEE International Conference on Robotics and Automation (ICRA) , 2000 , pp. 1194 - 1200 . doi:1 0 . 1 1 0 9 / R O B O T . 2 0 0 0 . 8 4 4 7 6 1 .

[4] D. B. Gennery , Traversability analysis and path planning for a planetary rover , Autonomous Robots 6 ( 1999 ) 131 - 146 . doi:1 0 . 1 0 2 3 / A : 1 0 0 8 8 3 1 4 2 6 9 6 6 .

[5] B. Rothrock , R. Kennedy , C. Cunningham , J. Papon , M. Heverly , M. Ono, SPOC: Deep Learningbased Terrain Classification for Mars Rover Missions , in: AIAA SPACE, 2016 , pp. 1 - 12 . doi:1 0 . 2 5 1 4 / 6 . 2 0 1 6 - 5 5 3 9 .

[6] B. Cafaro , M. Gianni , F. Pirri , M. Ruiz , A. Sinha , Terrain traversability in rescue environments , in: IEEE Safety Security and Rescue Robotics (SSRR) , 2013 , pp. 1 - 8 . doi:1 0 . 1 1 0 9 / S S R R . 2 0 1 3 . 6 7 1 9 3 5 8 .

[7] A. Huertas , L. Matthies , A. Rankin , Stereo-based tree traversability analysis for autonomous of-road navigation , in: IEEE Workshops on Applications of Computer Vision (WACV/MOTION), 2005 , pp. 210 - 217 . doi:1 0 . 1 1 0 9 / A C V M O T . 2 0 0 5 . 1 1 1 .

[8] M. A. Bekhti , Traversability Cost Prediction of Outdoor Terrains for Mobile Robot Using Image Features , Ph.D. thesis , Shizuoka University, 2020 .

[9] N. Hirose , A. Sadeghian , F. Xia , R. Martín-Martín , S. Savarese, VUNet: Dynamic scene view synthesis for traversability estimation using an rgb camera , Robotics and Automation Letters 4 ( 2019 ) 2062 - 2069 . doi:1 0 . 1 1 0 9 / L R A . 2 0 1 9 . 2 8 9 4 8 6 9 .

[10] A. Creswell , T. White , V. Dumoulin , K. Arulkumaran , B. Sengupta , A. A. Bharath , Generative adversarial networks: An overview , IEEE Signal Processing Magazine 35 ( 2018 ) 53 - 65 . doi:1 0 . 1 1 0 9 / M S P . 2 0 1 7 . 2 7 6 5 2 0 2 .

[11] K. Gopalakrishnan , S. K. Khaitan , A. Choudhary , A. Agrawal , Deep convolutional neural networks with transfer learning for computer vision-based data-driven pavement distress detection , Construction and Building Materials 157 ( 2017 ) 322 - 330 . doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . c o n b u i l d m a t . 2 0 1 7 . 0 9 . 1 1 0 .

[12] H.-C. Shin , H. R. Roth , M. Gao , L. Lu , Z. Xu , I. Nogues , J. Yao , D. Mollura , R. M. Summers , Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning , IEEE Transactions on Medical Imaging 35 ( 2016 ) 1285 - 1298 . doi:1 0 . 1 1 0 9 / T M I . 2 0 1 6 . 2 5 2 8 1 6 2 .

[13] R. Girshick , J. Donahue , T. Darrell , J. Malik , Regionbased convolutional networks for accurate object detection and segmentation , IEEE Transactions on Pattern Analysis and Machine Intelligence 38 ( 2016 ) 142 - 158 . doi:1 0 . 1 1 0 9 / T P A M I . 2 0 1 5 . 2 4 3 7 3 8 4 .

[14] R. Ribani , M. Marengoni , A survey of transfer learning for convolutional neural networks , in: Conference on Graphics, Patterns and Images Tutorials (SIBGRAPI-T) , 2019 , pp. 47 - 57 . doi:1 0 . 1 1 0 9 / S I B G R A P I - T . 2 0 1 9 . 0 0 0 1 0 .

[15] M. E. Taylor , S. Whiteson, P. Stone , Transfer via inter-task mappings in policy search reinforcement learning , in: International Conference on Autonomous Agents and Multiagent Systems (AAMAS) , 2007 , pp. 156 - 163 . doi: 10 .1145/1329125. 1329170.

[16] N. Makondo , M. Hiratsuka , B. Rosman , , O. Hasegawa , A non-linear manifold alignment approach to robot learning from demonstrations , Journal of Robotics and Mechatronics 30 ( 2018 ) 265 - 281 . doi: 10 .20965/jrm. 2018 .p0265.

[17] K. Ogawa , P. Hartono , Infusing common-sensical prior knowledge into topological representations of learning robots , Artificial Life and Robotics ( 2022 ) 1 - 10 . doi: 10 .1007/S10015- 022- 00776- 5.

[18] C. Devin , A. Gupta , T. Darrell , P. Abbeel , S. Levine , Learning modular neural network policies for multitask and multi-robot transfer , in: IEEE International Conference on Robotics and Automation (ICRA) , 2017 , pp. 2169 - 2176 . doi: 10 .1109/ICRA. 2017 . 7989250 .

[19] P. Arena , C. F. Blanco , A. Li Noce , S. Tafara , L. Patanè , Learning traversability map of different robotic platforms for unstructured terrains path planning , in: International Joint Conference on Neural Networks (IJCNN) , 2020 , pp. 1 - 8 . doi: 10 .1109/IJCNN48605. 2020 . 9207423 .

[20] J. Zelinka , M. Prágr , R. Szadkowski , J. Bayer , J. Faigl , Traversability transfer learning between robots with diferent cost assessment policies , in: 2021 Modelling and Simulation for Autonomous Systems (MESAS) , 2022 , pp. 333 - 344 . doi: 10 .1007/ 978- 3- 030 - 98260- 7_ 21 .

[21] J. Bayer , J. Faigl , Decentralized topological mapping for multi-robot autonomous exploration under low-bandwidth communication , in: European Conference on Mobile Robots (ECMR) , 2021 , pp. 1 - 7 . doi: 10 .1109/ECMR50962. 2021 . 9568824 .

[22] M. Forouhar , P. Čížek , J. Faigl, SCARAB II : A small versatile six-legged walking robot , in: 5th Full-Day Workshop on Legged Robots at IEEE International Conference on Robotics and Automation (ICRA) , 2021 , pp. 1 - 2 .