2019 International Conference on Robotics and Automation (ICRA)

Palais des congres de Montreal, Montreal, Canada, May 20-24, 2019

Single-shot Foothold Selection and Constraint Evaluation for Quadruped

Locomotion

Dominik Belter1 , Jakub Bednarek1 , Hsiu-Chin Lin2 , Guiyang Xin2 , Michael Mistry2

Abstract— In this paper, we propose a method for select-

ing the optimal footholds for legged systems. The goal of
the proposed method is to find the best foothold for the
swing leg on a local elevation map. First, we evaluate the
geometrical characteristics of each cell on the elevation map,
checks kinematic constraints and collisions. Then, we apply
the Convolutional Neural Network to learn the relationship
between the local elevation map and the quality of potential
footholds. During execution time, the controller obtains the
qualitative measurement of each potential foothold from the
neural model. This method evaluates hundreds of potential
footholds and checks multiple constraints in a single step
which takes 10 ms on a standard computer without GPU.
The experiments were carried out on a quadruped robot
walking over rough terrain in both simulation and real robotic Fig. 1. The foothold selection problem for a quadruped robot. The
platforms. region of the elevation map below the i -th leg is evaluated to find the
best foothold. Each candidate position of the foot has to be kinematically
I. I NTRODUCTION feasible and collision-free.

Locomotion in challenging terrain requires careful selec-

tion of footholds. The robot should select stable support
map sufficiently represents the terrain, guarantees quick
for each foot to avoid slippages and falls. This strategy is
access to each cell, and can be directly transformed to
crucial when the robot needs to deal with highly irregular
grayscale image to feed the Convolutional Neural Net-
terrain. A challenging example is maneuvering in extreme
work [8] which we use in this research.
environments such as a mine, or the aftermath of a natural
The problem of foothold selection is presented in Fig. 1.
disaster where the robot can find only a few acceptable
The robot evaluates the region (elevation map) to select
footholds. A poor foothold selection method means that
the best foothold that fulfills a set of kinematics (e.g.,
the robot may fall and cannot execute the mission.
workspace limit, self-collision, etc) and environmental
In contrast, other types of locomotion assume that the
constraints (e.g., avoiding sharp edges). In the classical
robot walks dynamically on rough terrain and stabilizes
approach, all constraints are verified sequentially by the
its posture using fast control algorithms and compliant
controller of the robot during walking [9]. On the other
legs [1], [2]. Stable locomotion relies on the capability of
hand, online optimization is time-consuming and not
the controller to cope with disturbances (e.g. slippages)
feasible for real-time control.
resulting from unstable footholds. However, this approach
In this research, we propose a computationally efficient
is only efficient on moderately rough terrain and it will fail
solution for foothold selection. We applied a neural net-
when the robot has to face extreme environments such as
work to learn a model (off-line) that maps the properties
a mine, or the aftermath of a natural disaster.
of the terrain to the quality of a potential foothold while
To efficiently navigate in a more extreme environment,
excluding footholds which are risky or kinematically in-
perception systems are required to detect high obstacles
feasible. During execution time, we efficiently predict the
so the robot can avoid them while walking [3]. A full
quality of a potential foothold from the learned model.
3D model of the environment can be obtained using
The proposed method is verified on a quadruped robot
terrain mapping methods, such as OctoMap [4], Normal
walking over rough terrain, in both simulation and real
Distribution Transform Occupancy Maps (NDT-OM) [5], or
robot platforms.
elevation map [6], [7]. Among all of the above, elevation

*This research was supported by EU Horizon 2020 project THING. This

II. R ELATED W ORK
work has been conducted as part of ANYmal Research, a community
to advance legged robotics. We gratefully acknowledge the support of
The problem of foothold selection is similar to the prob-
NVIDIA Corporation with the donation of the Titan Xp GPU used for lem of multi-finger grasping and was studied widely by
this research. the robotics community. Recent development in this field
1 Institute of Control, Robotics and Information Engi-
includes the method which use local geometrical prop-
neering, Poznan University of Technology, Poznan, Poland
[email protected] erties of the objects to find the acceptable positions of
2 School of Informatics, University of Edinburgh, Edinburgh, UK the fingertips on the object’s surface [10]. The grasp con-

978-1-5386-6027-0/19/$31.00 ©2019 IEEE 7441

figurations are trained from real examples. The collision
and kinematic constraints are taken into account during
W
the inference procedure. Recently, deep neural network,
such as the Convolutional Neural Networks (CNN) gained
high popularity in robotics applications. In grasp, the Li
CNN is applied to select feasible grasp and robotics finger
positions on the object’s surface using point cloud [11] or
depth images [12].
Most approaches for the foothold selection are based
Mi
on the local features computed for the terrain surface,
such as the inclination of the terrain, roughness, and local Ni
curvature from the elevation maps [13]. These features
are provided to the input of the simple neural network
which was trained on the data provided by human experts. Fig. 2. Center of the local map M i used for foothold selection is located
below the i -th leg joint L i . The foothold selection algorithm considers
Another approach takes the elevation map and estimates also the nominal position of the foot Ni .
a probability map that is related to the capability of each
cell to provide stable support for the robot’s feet [14].
The StarlETH robot is equipped with the haptic device of rough terrain and a reference foothold scorer [9]. The
on the feet, which explores and evaluates the potential learnt network implicitly stores information about the
footholds without human supervision [15]. The HyQ robot kinematic and collision model of the leg and rejects
focuses more on the reflexes which stabilize the robot [16], footholds which are outside the workspace. The main
but it also uses visual information about the terrain to contributions of this paper include the following:
avoid risky footholds [17]. The robot corrects the nominal
1) We are the first who show that the CNN can be used
foothold positions according to the output from the visual
to evaluate geometrical properties of the potential
pattern classifier applied on the terrain patches.
footholds and simultaneously consider all kinematic
Great progress in the field of autonomous legged lo-
constraints which are related to the model of the
comotion on rough terrain was done on the quadruped
robot. Comparing to formal work which optimizes
robot LittleDog [18]. The authors proposed a terrain scorer
constraints on-line, this significantly reduces the
which computes the spatial relationship between a con-
online computational time.
sidered point and its neighboring points and then rejects
2) A transfer learning approach that learns a model
points which are located on edges, large slope, the base
from data gathered in the simulation. With this ap-
of a cliff, or inside of a hole.
proach, we provide a sufficient number of examples
A learning-based method was proposed to evaluate
to train a neural network without the need for using
terrain templates based on the human demonstration [19].
the real robot. We also show that the obtained neural
The terrain scorer approach is also adapted in [20],
model is successfully transferred to the real robot.
where the weights of geometric features of the terrain
are obtained during training and then used for the foot- III. F OOTHOLD S ELECTION M ODULE
steps planning. Very recently, the CNN classifier for the
footholds has been proposed [21]. We propose a foothold selection module to evaluate po-
The foothold selection method for a six-legged robot is tential footholds which are inside the local map extracted
represented by the method implemented on the Lauron IV from the global elevation map.
robot [22]. The foothold selection module considers points
A. Problem Definition
around initial foothold and takes into account elevation
credibility, the mean height, and the height variance of A global map is an elevation map built online by the robot
the cells. The six-legged Messor robot learns which points and aligned with the world coordinate system W , where
on the elevation map can provide stable support from the center of this map is the center of the robot projected
simulation data [9]. Then, the trained Gaussian Mixture on the ground. The size of the global map is 6 × 6m and
is used to select the footholds in the RRT-based motion the size of each cell is 2 × 2 cm.
planner [23]. The kinematic and self-collision constraints A local map is a subset of the global elevation map
are also taken into account. However, this process signif- where the center M i is a point below the hip joint of the
icantly slow-downs the foothold selection process. considered leg (presented in Fig. 2). The size of the local
map, which is 40×40 cells, covers the kinematic range of
A. Approach and Contribution the leg.
In this paper, we propose a novel method to evaluate A nominal foothold Ni is the desired position of the
potential footholds for the quadruped robot in a single foot for a given step length and assuming that the robot
step using CNN. We collect data for training the network is walking on flat terrain. Lastly, a potential foothold is a
using the kinematic model of the robot, elevation map cell inside of the local map.

7442
 Our goal is to select an optimal foothold from a set
of potential footholds online using the information from
the local map that satisfies the following constraints.
0.4 z [m]
First, the robot should avoid selecting footholds on sharp 0.1
-6
edges and/or steep slopes because they are potentially -0.2
6 -4

risky. Second, the selected foothold should be within the 4 -2

2 0
kinematic limit of the robot (inside the workspace of the x [m] 0
y [m]
2
leg). Also, the robot should avoid self-collisions and check -2
-4 4

whether the thigh or shank collides with the terrain. -6 6

B. Dataset Fig. 3. Elevation map created offline to collect data for training the
neural network.
Training a model from the robot is expensive, time-
consuming, and dangerous since sensor data is prone to
2) terrain cost: The terrain cost c m evaluates the prop-
noise. Thus, we took the transfer learning approach, where
erty of the local map. We use the same cost function
data are gathered in simulation but used on the real robot.
with the hexapod Messor robot [9] since both robots
To train the neural network we collect the samples on
have a similar hemispherical foot. The c m value is
the 12 × 12m elevation map presented in Fig. 3. The map
normalized.
was created offline by composing maps obtained during
various experiments on the robot. We also added the flat Finally, we compute the final cost of the considered
region, steps with various height, concavities, and bumps foothold c f and scale the cost to the range [0, 255]:
to increase the variation of foothold examples. ck + 2 · cm
cf =· 255. (1)
Since the robot is symmetrical, we train the models for 3
the right legs and adapt it for the left legs. To this end, For each cell, we repeat the above procedure and save
we have to flip horizontally the input terrain map and the input (elevation map) and the output (terrain cost).
after inference, we flip horizontally the obtained cost map. We collected 20000 training pairs for each leg.
Therefore, we only need to collect data for two legs and Examples of training data are presented in Fig. 4. The
train two separate models. first two columns present the computed cost maps for
To generate training data, we randomly select the po- the flat terrain. In this case, the output depends on the
sition of the robot on the map (horizontal position and leg’s workspace and the kinematic margin. The distance
distance to the ground). The orientation of the robot on between the terrain and the robot is larger on the map in
the horizontal plane (yaw angle) is randomly selected Fig. 4a than in Fig. 4b. The output cost map also differs.
from four main orientations: n · Π2 , for n = 0, 1, 2, 3. For the The obtained cost maps (Fig. 4f and Fig. 4g) represents the
obtained pose of the robot, we compute the pose of the horizontal cross-section over the workspace of the robot’s
i -th leg and extract a 40×40 local map. For each cell of the leg. The yellow cells represent positions of the foot which
map, we quality the cost of a foothold c f ∈ [0, 255] based are outside the workspace and are inaccessible for the
on a set of constraints. This quantity is first evaluated robot (c f = 255). Collisions, the edges on the obstacles or
based on the following hard constraints: slopes also increase significantly the cost of footholds and
1) kinematic range of the leg: If the given position of the neural network classifies them as inaccessible (yellow
the foot is outside the workspace of the considered color). The red cells correspond to acceptable footholds. In
leg, we set the cost to the maximal value c f = 255, the following examples in Fig. 4c–e the terrain is irregular
and we do not check other constraints. and we can observe how the workspace of the robot is
2) self-collisions and collision with the ground: We use limited by the terrain shape.
Flexible Collision Library [24] to determine whether C. Convolutional Neural Network
there is any collision between any pairs of rigid
We aim to learn a mapping between the features extracted
bodies and with the ground, except the collision
from a local elevation map and the quality of potential
between the foot and the terrain. A foothold is
footholds. In this work, we choose Convolutional Neural
rejected (i.e., c f = 255) if there is a collision, and
Network (CNN), since this architecture runs in real-time
we do not check further.
on machines without GPUs.
If a given potential foothold is collision-free and the foot is Because CNN is much more efficient in solving classifi-
within leg’s workspace, the quality of selecting a foothold cation than regression, we discretize the terrain costs into
is evaluated using the following evaluation criteria: C different classes. In this work, C is set to 14 since it is
1) kinematic margin: The kinematic margin c k is the sufficient to distinguish between weak and good footholds
distance between the current position of the foot and we can easily provide a sufficient number of training
and the border of the workspace. The maximal value samples for each class.
of c k means that the leg has the maximal motion The proposed CNN architecture is an Efficient Residual
range. The c k value is normalized. Factorized ConvNet (ERF) first introduced in [25]. The

7443
 a b c d e

f g h i j

Fig. 4. Example training data: local elevation maps (a,b,c,d,e), and corresponding terrain cost (f,g,h,i,j) (red color – acceptable footholds, yellow
– unacceptable footholds). Note that both subfigures a and b represent flat terrain but elevation is different. Thus, the acceptable region (red area
which corresponds to the leg’s workspace) obtained from the neural network is different in subfigures f and g.

Leg Accuracy [%] IoU

front leg 82.61 49.9
rear leg 82.61 49.88
TABLE I
A CCURACY AND I NTERSECTION OVER U NION (I O U) OBTAINED ON
VALIDATION SET FOR FRONT AND REAR LEG MODELS

128
cross-entropy loss and regularization loss. In the training
64 64
Input
16 16
Output
dataset, the most examples are provided for the class
40x40x1 40x40xC
which represents footholds inaccessible for the robot. To
Fig. 5. Model of the ERF network. Light blue blocks represent handle unbalanced data, the cross-entropy is additionally
downsampling, dark blue - upsampling by transposed convolution and weighted [27] based on the frequency of occurrence.
white blocks show residual layers. Numbers below blocks describes the Namely, the weight of the i t h class w i is defined by
number of feature maps used in specific levels. C denotes the number
of classes (14 in the current implementation). 1
wi = (2)
log(c + p i )
where c = 1.08 is a constant and p i is a probability of the
characteristics of this model is the modification of the
occurrence of the i t h class in the entire training dataset.
residual layer [26] called residual non-bottleneck 1D layer.
We use the method presented in [28] for training the
The 2D convolution with a 3×3 filter is replaced by two
model with an initial learning rate of 5e-4. Additionally,
2D convolutions with filters shape 3×1 and 1×3. This
the exponential decay was applied after each epoch to the
approach reduces the number of variables and complexity.
learning rate with a factor of 0.98. Because of the nature
The ERF model is shown in Fig. 5. First, the input of the training examples, we can’t use any of the known
data is processed twice by downsampling blocks. The data augmentation methods.
downsampling blocks are created from the concatenation In order to measure the quality of models accuracy, an
of the max pooling and 2D convolution with a 3×3 Intersection over Union (IoU) metrics were calculated. The
filter and a stride of 2. The concatenation is followed learning process took place in 500 epochs. The results
by the activation function. Then, five residual layers and obtained by two ERF models for front and rear legs are
another downsample block are added. The output of the shown in Tab. I. The IoU value is higher than 82%. Note
encoder part is processed by eight residual layers which that, this value does not represent directly the quality of
are interwoven with different dilation rate applied to the the foothold selection module. Although the learnt model
convolutions. The decoder part of the model consists of misclassifies 22% of the footholds, most of the errors
two series of convolutional upsampling and two residual are between neighboring classes, which is not a crucial
layers. The upsampling is performed by transposing con- problem. For example, if the foothold is classified as a
volution with a stride of 2. The output of the model is class number 13 instead of the class number 14 it is still
produced by upsampling convolution with 2×2 filters and considered as a very weak foothold.
strides of 2, where the number of filters is equal to the
number of classes. The Activation function used in each D. Inference procedure
nonlinear layer is a rectified linear unit (ReLU). The inference procedure is presented in Fig. 7. In the
The optimized objective of the model is composed of first step, we get submap from the global map built by

7444
 a b

Fig. 6. Environment configuration during experiment with the ANYmal robot on stairs (a) and on rough terrain (b) in the Gazebo simulator. Blue
lines represent feet trajectories.

Extract local Convert to the

the robot. The obtained map is aligned with the world elevation Map normalized image
coordinate system W but our neural network uses the
elevation map which is aligned with the robot coordinate
system. Thus, we rotate the obtained local map by the
current orientation of the robot on the horizontal plane
(yaw angle). Some information about cells at the corners
Compute the
is lost during this rotation, therefore, we take a slightly
distance to the CNN inference
larger map for rotation purpose. Before rotation, the size
nominal foothold
of the local map is 51×51 cells and after rotation, we crop
the map to size 40×40 cells.
In the next step, we convert the obtained elevation
map to the image. To this end, we compute the distances
between the i -th leg coordinate system L i and each cell
of the map. We use 8-bit grayscale images as an input Find the minimal cost
to the network so the obtained distance values are fitted
into range 0–255. We use a constant normalization factor
(0.85 m) for each leg of the robot. The obtained image
which represents the terrain patch around the consider
leg is the input to the neural network model. Fig. 7. Foothold selection procedure on the local elevation map
The network classifies each pixel on the image, and
the cost at each pixel corresponds to the cost of taking
that foothold. The example inference results for the input IV. R ESULTS
image representing stairs are presented in Fig. 7. The First experiments are performed in the Gazebo simulator.
pixels which are located on the edges between steps on We verified the proposed foothold selection method on
the output image are brighter which means that the robot the ANYmal robot walking on stairs (Fig. 6a) and on
should avoid these footholds. At the same stage of the rough terrain (Fig. 6b). The robot uses a simulated Intel
inference procedure, we compute the distance from the RealSense D435 RGB-D sensor to build a map of the
nominal foothold d n . Then, we compute the final cost environment [7]. We use the controller presented in [29]
c final for each pixel (foothold): to plan the foot trajectories above the obstacles, estimate
the state of the robot and execute planned trajectories.
We only replaced the foothold selection model.
c final = c f + k · d n , (3)
The example inference results are presented in Fig. 8.
We provide the terrain patches extracted from the global
where c f is the cost computed by the neural network and elevation map, the distance between potential footholds
k is the constant value which determines the influence of and the nominal foothold, and the output from the CNN.
the distance from the nominal foothold on the final cost It is clearly visible from the result obtained on the stairs
of the potential foothold. In the experiments presented in that the robot avoids placing its feet on the edges. These
the paper the k value is set to 140. We compute the final regions are classified by the neural network as risky
cost c final for each pixel on the image (c.f. Fig. 7) and we and rejected by the foothold selection module. Similar
find the minimal value. Then, the pixel with the minimal behavior can be observed in the results obtained on rough
cost in image coordinates is converted into the 3D point terrain. In this case, the obstacles are more irregular. For
in the world coordinate system. The obtained value is sent both patches obtained on rough terrain, the region in the
to the controller which executes the motion for the given center of the workspace has a similar cost. In this case, the
foothold. distance from the nominal foothold plays an important

7445
 a a

b
d

e c

Terrain patch Distance from CNN output Selected

the nominal red color - low cost foothold
foothold
Fig. 9. Experiment with the ANYmal robot on the rough terrain mockup
(a): example terrain patches (b) and CNN output (c)
Fig. 8. Example inference results obtained during the experiment on
stairs (a,b,c) and on rough terrain (d,e). The stairs and rough terrain are
presented in Fig. 6a and Fig. 6b, respectively.
In contrast to previous work based on manually com-
puted features of the terrain [13], the method based
role. The selected foothold is close to the nominal foothold on CNN extracts features automatically from data. Our
but still on the position with acceptable foothold cost method evaluates 1600 potential footholds and checks
predicted by the CNN. constraints in a single inference step. In contrast to pre-
We also compare our method with the foothold se- vious work, our new approach only needs approximately
lection method proposed in [29]. The average foothold 10 ms on the CPU to infer from the input elevation map.
selection time for [29] is 3.44 ms per 10 samples, while Finally, we performed the experiments on the real robot
our method produces the output in 151.93 ms (the CNN walking over a customized rough terrain consists of 12
inference takes 10 ms on the Intel CPU i7-2640M and blocks with different slopes and orientation. The robot
the rest of the time is consumed by getting data from should avoid stepping on the tips or the edges. The
the global elevation map and preparing input data for example results are presented in Fig. 9. The obtained
the CNN). Although our method may not be the most elevation map (Fig. 9b) is less accurate than the map
time efficient one, we outperform the previous work in obtained in the simulation experiments due to noise, but
the choice of footholds. We validated both methods on a the robot can still identify risky edges and place its feet
stair with 8 steps, each step is 0.18 m high and 0.29 wide. on the stable positions (see supplementary video).
In average, the robot reached 5 steps (in 10 trials) using
our method while the robot failed after 2.6 steps using V. C ONCLUSIONS AND F UTURE W ORK
method from [29]. Note that, the failures in both methods
are not caused by improper foothold selection but by the In this paper, we propose a novel foothold selection
stability controller. In all unsuccessful cases, the robot method for legged systems. In contrast to methods known
fails because of the lack of stability. Our method performs from the literature, the proposed method learns a model
better because it takes into account the workspace of the that evaluates the terrain patches and checks all con-
leg and selects footholds which are far from the border of straints in a single step. The time complexity for the
the leg’s workspace. As a result, the posture of the robot inference is significantly reduced. With the proposed
is more stable during walking. method, the robot avoids placing its feet on the edges or
To demonstrate the benefit of learning a model instead steep slopes. The neural network also implicitly takes into
of optimizing all constraints online, we compared our account the kinematic range of the leg and detects self-
method with the method proposed in [9]. We applied the collisions and collisions with the ground. The proposed
method from [9] to the ANYmal robot and we evaluated foothold selection module is integrated with the controller
the same constraints. Instead of using CNN we use the of the robot. In the simulation and experiments with the
kinematic model of the robot to check the workspace real robot, we present the properties and the efficiency of
of the legs and we use Flexible Collision Library [24] to the proposed method.
detect collisions. In this case, the foothold evaluation in In the future, we plan to use the neural network to
the 40×40 window takes 2008.81 ms which is more than optimize simultaneously the foothold position and the
10 times slower than the proposed application of CNN. posture of the robot.

7446
 R EFERENCES [15] M. A. Hoepflinger, M. Hutter, C. Gehring, M. Bloesch, and R.
Siegwart, Unsupervised identification and prediction of foothold
[1] M. Hutter, C. Gehring, D. Jud, A. Lauber, C. Bellicoso, V. Tsounis,
robustness, IEEE International Conference on Robotics and Au-
J. Hwangbo, K. Bodie, P. Fánkhauser, M. Bloesch, R. Diethelm,
tomation, pp. 3293–3298, 2013
S. Bachmann, A. Melzer, M. Hoepflinger, ANY-mal - a highly
[16] M. Focchi, R. Orsolino, M. Camurri, V. Barasuol, C. Mastalli, D.G.
mobile and dynamic quadrupedal robot, Proc. IEEE International
Caldwell, C. Semini, Heuristic Planning for Rough Terrain Locomo-
Conference on Intelligent Robots, pp. 38–44, 2016
tion in Presence of External Disturbances and Variable Perception
[2] D. Hyun, D. Jin, S. Seok, J. Lee, S. Kim, High speed trot running:
Quality, arXiv:1805.10238v3, 2018
Implementation of a hierarchical controller using proprioceptive
[17] V. Barasuol, M. Camurri, S. Bazeille, D.G. Caldwell, C. Semini,
impedance control on the MIT Cheetah, The International Journal
Reactive Trotting with Foot Placement Corrections through Visual
of Robotics Research, Vol. 33(11), pp. 1417–1445, 2014
Pattern Classification, IEEE/RSJ International Conference on Intel-
[3] D. Wooden, M. Malchano, K. Blankespoor, A. Howardy, A.A. Rizzi,
ligent Robots and Systems, pp. 5734–5741, 2015
M. Raibert, Autonomous navigation for BigDog, 2010 IEEE Inter-
[18] J. Rebula, P. Neuhaus, B. Bonnlander, M. Johnson, and J. Pratt. A
national Conference on Robotics and Automation, pp. 4736–4741,
controller for the LittleDog quadruped walking on rough terrain. In
2010
Proc. IEEE Int. Conf. on Robotics and Automation, pp. 1467–1473,
[4] A. Hornung, K. Wurm, M. Bennewitz, C. Stachniss, W. Burgard,
2007
OctoMap: An efficient probabilistic 3D mapping framework based
[19] M. Kalakrishnan, J. Buchli, P. Pastor, S. Schaal, Learning locomotion
on octrees, Autonomous Robots, Vol. 34(3), pp. 189–206, 2013
over rough terrain using terrain templates. In Proc. IEEE Int. Conf.
[5] J. Saarinen, H. Andreasson, T. Stoyanov, A.J. Lilienthal, 3D normal
on Intelligent Robots and Systems, pp. 167–172, 2009
distributions transform occupancy maps: An efficient representa-
[20] J. Kolter, M. Rodgers, A. Ng, A control architecture for quadruped
tion for mapping in dynamic environments, International Journal
locomotion over rough terrain. In Proc. IEEE Int. Conf. on Robotics
of Robotics Research, Vol. 32(14), pp. 1627–1644, 2013
and Automation, pp. 811–818, 2008
[6] D. Belter, P. Łabecki, Fánkhauser, R. Siegwart, RGB-D terrain percep- [21] O. Villarreal, V. Barasuol, M. Camurri, M. Focchi, L. Franceschi,
tion and dense mapping for legged robots, International Journal of M. Pontil, D.G. Caldwell, C. Semini, Fast and Continuous Foothold
Applied Mathematics and Computer Science, Vol. 26(1), pp. 81–97, Adaptation for Dynamic Locomotion through Convolutional Neural
2016Automation, pp. 5761–5768, 2018 Networks, arXiv, 2019
[7] P. Fánkhauser, M. Bloesch, C. Gehring, M. Hutter, R. Siegwart, [22] A. Roennau, T. Kerscher, M. Ziegenmeyer, J. M. Zöllner, and R.
Robot-Centric Elevation Mapping with Uncertainty Estimates, Inter- Dillmann. Six-legged walking in rough terrain based on foot point
national Conference on Climbing and Walking Robots, pp. 433–440, planning. In O. Tosun, M. Tokhi, G. Virk, and H.L. Akin (Eds.), Mo-
2014 bile Robotics: Solutions and Challenges, World Scientific, pp. 591–
[8] A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet classification 598, 2009
with deep convolutional neural networks, Proc. of the International [23] D. Belter, P. Łabecki, P. Skrzypczyński, Adaptive Motion Planning for
Conference on Neural Information Processing Systems, pp. 1097– Autonomous Rough Terrain Traversal with a Walking Robot, Journal
1105, 2012 of Field Robotics, Vol. 33(3), pp. 337–370, 2016
[9] D. Belter, P. Skrzypczyśki, Rough terrain mapping and classification [24] J. Pan, S. Chitta, D. Manocha, FCL: A General Purpose Library for
for foothold selection in a walking robot, Journal of Field Robotics, Collision and Proximity Queries, IEEE International Conference on
Vol. 28(4), pp. 497-528, 2011 Robotics and Automation, pp. 3859–3866, 2012
[10] M. Kopicki, R. Detry, M. Adjigble, R. Stolkin, A. Leonardis, J.L. Wyatt, [25] E. Romera, J.M. Alvarez, L.M. Bergasa, R. Arroy, ERFNet: Efficient
One shot learning and generation of dexterous grasps for novel Residual Factorized ConvNet for Real-Time Semantic Segmentation.
objects, International Journal of Robotics Research, pp. 959–976, IEEE Transactions on Intelligent Transportation Systems Vol. 19,
Vol. 35(8), 2015 pp. 263–272, 2018
[11] J. Mahler, J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. Aparicio, [26] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image
K. Goldberg, Dex-Net 2.0: Deep Learning to Plan Robust Grasps Recognition, IEEE Conference on Computer Vision and Pattern
with Synthetic Point Clouds and Analytic Grasp Metrics, Robotics Recognition, pp. 837–840, 2017
Science and Systems, 2017 [27] A. Paszke, A. Chaurasia, S. Kim, E. Culurciello, ENet: A Deep
[12] M. Gualtieri, A. ten Pas, K. Saenko, R. Platt, High precision grasp Neural Network Architecture for Real-Time Semantic Segmentation,
pose detection in dense clutter, IEEE/RSJ International Conference arXiv:1606.02147, 2016
on Inteligent Robots and Systems, pp. 598–605, 2016 [28] D.P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization,
[13] E. Krotkov, R. Simmons. Perception, planning, and control for International Conference on Learning Representations, 2014
autonomous walking with the Ambler planetary rover. International [29] P. Fankhauser, M. Bjelonic, C.D. Bellicoso, T. Miki, M. Hutter,
Journal of Robotics Research, Vol. 15(2), pp. 155–180, 1996 Robust Rough-Terrain Locomotion with a Quadrupedal Robot, IEEE
[14] C.-H. Chen, V. Kumar. Motion planning of walking robots in International Conference on Robotics and and Systems, Daejeon,
environments with uncertainty. In Proc. IEEE Int. Conf. on Robotics South Korea, pp. 38–44, 2016
and Automation, pp. 3277–3282, 2009

7447