You are on page 1of 7

Research Collection

Student Paper

Hybrid Contact Detection and Force Estimation during Compliant


Manipulation

Author(s):
Wang, Tianlu

Publication Date:
2021

Permanent Link:
https://doi.org/10.3929/ethz-b-000463448

Rights / License:
In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For more
information please consult the Terms of use.

ETH Library
Hybrid Contact Detection and Force Estimation during Compliant
Manipulation
Author: Tianlu Wang
Supervisors: Martin Wermelinger, Jan Carius, and Prof. Marco Hutter

Abstract— This work introduces a hybrid framework of


sensorless contact detection and external force estimation for
robotic arms with general manipulation tasks. Firstly, the fast
contact detection based on conditional probability, which is
used to observe the existence of contact, is derived. Secondly,
the method of accurate external force estimation based on
time-varying signal tracking is illustrated. Combining those
two methods, a hybrid algorithm is formulated and experi-
mentally tested on ANYpulator, a 6 DoF robot arm composed
of series elastic actuators. The accuracy, precision, and time
response of the proposed algorithm are analyzed in detail and
evaluated with benchmark measurements provided by a 6-axis
force/torque sensor. Two force tracking tasks are shown to
further verify the effectiveness of the algorithm.

I. INTRODUCTION
Contact detection and external force estimation are essen- Fig. 1. The 6 DoF robot arm ANYpulator, composed of series elastic
tial for robots during manipulation and locomotion tasks. actuators, is used for the experimental verification. At the end-effector there
is a 6 DoF force sensor, used only as a reference benchmark.
Reliable feedback of contact force amplitude and direc-
tion are important information for control and planning.
Furthermore, fast detection of external forces can assure
robot arms. The estimator has been extended to a second-
safe interaction and effective collaboration with objects and
order one to further eliminate the noise, and several appli-
humans nearby. For traditional industrial robots and most
cations like hand-guided teaching have been demonstrated
commercial robots, contact detection and force estimation
[2], [3]. Another trend is to use a probabilistic framework
by force/torque sensors are still mostly applied, and various
to regard contact detection as a classification problem, that
commercial products are available on the market. However,
is, contact or no contact. This concept is quite recent, and
limited by space, the installation of sensors might not be
most applications are towards gait phases switching for
applicable for all robots. Furthermore, sensor readings are
quadrupedal robots, like ANYmal [4] and HyQ [5].
often prone to drift caused by sharp temperature or humidity
variation in manipulation environments. For convenience of illustration, we summarize relevant
concepts into detection and estimation. Contact detection
To avoid such sensors but still give robots a way to
observes the existence of contact, and estimation is aimed
perceive contacts, the concept of sensorless contact detection
at external force approximation. In former works, the pre-
and force estimation was introduced by De Luca et.al. [1]
cision, accuracy, and time delay issues of estimation have
for manipulators. Here, ‘sensorless’ refers to the fact that
not been analyzed in detail properly, compared with force
there is no extra force/torque sensor needed at the point of
sensor readings, for example (precision here refers to the
contact to measure the external wrench (force and torque),
degree of noise, and accuracy refers to the deviation from
although measurements of joint torques are still required.
the benchmark values). Moreover, while the residual-based
They designed a first-order low pass filter of a so-called
estimation provides a good approximation of the interaction
residual, which is the effect of external forces felt at the joint
force in a contact situation, it is prone to model errors and
level. This idea is widely used in recent applications with
introduces an inherent time delay. It is therefore not suitable
This work was supported in part by the Swiss National Science Founda- as a fast indicator to decide if a contact occurs during a
tion (SNF) and the National Centre of Competence in Research Digital dynamic motion. We therefore combine estimation through
Fabrication. The study was carried out as a semester thesis at Robotic residuals with a detection method using the probabilistic
Systems Lab (RSL), ETH Zurich.
Tianlu Wang is now with the Department of Information Technology and framework and use their complementing pieces of informa-
Electrical Engineering, ETH Zurich, Switzerland, and Max Planck Institute tion to ultimately estimate the contact state.
for Intelligent Systems, Stuttgart, Germany (email: tiawang@ethz.ch). Since both former works on a probabilistic framework
Martin Wermelinger, Jan Carius, and Marco Hutter are with Robotic
Systems Lab (RSL), ETH Zurich, Switzerland (email: {martiwer, are designed in particular for walking robots that touch the
jan.carius}@mavt.ethz.ch and mahutter@ethz.ch). ground cyclically, we generalize this formulation to imple-
ment our robust hybrid algorithm in the context of general contributions to create the new measurement update for the
robotic manipulators. Specifically, we make the following combined filter. The method has been tested successfully
contributions: on the quadrupedal HyQ. Unfortunately, the probabilistic
• Proposal and implementation of the hybrid algorithm, frameworks cannot provide estimations for the external force
which is composed of two parts, contact detection by (value approximation) yet.
conditional probability and contact force estimation by To the best of our knowledge, there is no former work on
time-varying signal tracking; probabilistic contact detection implemented and tested.
• Analysis of the algorithm’s precision, accuracy and time
response; III. M ETHOD
• Demonstration of force tracking experiments, showing The task of hybrid sensorless contact detection and force
the applicability of the proposed algorithm for manipu- estimation can be summarized into two sequential steps. The
lators. first step is to detect the existence of contact by a fast
response algorithm. Then, the force/torque amplitude and
II. R ELATED W ORK
direction from the contact is estimated by a signal tracking
Major related research on sensorless contact detection and algorithm. The two steps and the combined algorithm are
estimation focuses on the detection part. Methods can be illustrated in the following part. Our derivation is based on
categorized into residuals approach and probabilistic frame- the assumption that the external force is only applied to the
works. end-effector.
A. Residuals Approach A. Contact Detection through Conditional Probability
The most widely used ideas on sensorless contact detection The objective of contact detection is to compute the
originate from De Luca’s work on the DLR robotic arm [1]. probability of the contact state Xt based on the current
He introduces residuals, which are regarded as an estima- state of the robot. The contact state Xt is binary, Xt ∈
tor to observe the external force at the joint level. When {C, N }, where C and N represent ‘in contact’ and ‘not in
transferred to Laplacian frequency domain, the estimator can contact’, respectively. The probability of ‘contact’ then can
also be seen as a first-order low pass filter, which eliminates be expressed as conditional probability
some amounts of noise during the detection. Compared with
the direct calculation by the system dynamics equation, the P (Xt = C|q, q̇, q̈), (1)
formulation is motivated by the fact that joint accelerations
are not explicitly required. From the application side, the where q, q̇ and q̈ represent generalized coordinate, gen-
contact detection has been used to trigger fast collision eralized velocity and generalized acceleration of the robot
reaction, and the safe interaction between human and robot arm. The state of the robot is uniquely given by the pair
has also been demonstrated. {q, q̇}, which can be obtained from joint encoder readings.
In the recent two years, the applications of residuals are By Bayes’ rule, we expand the conditional probability to
further explored. In Sang-Duck’s work [2], a force guiding P (q, q̇, q̈|Xt )P (Xt )
controller based on residual computation has been verified P (Xt |q, q̇, q̈) = P . (2)
Xt ∈{C,N } P (q, q̇, q̈|Xt )P (Xt )
on a 6 DoF arm.
Through conditioning and chain rule, the term P (q, q̇, q̈|Xt )
B. Probabilistic Framework
can be further represented as
The idea of probabilistic contact detection originates from
the problem of force sensor’s unreliability on fast contact P (q, q̇, q̈|Xt ) = P (q̈|Xt , q, q̇) P (q̇|Xt , q) P (q|Xt ) . (3)
| {z } | {z } | {z }
detection on quadrupedal robots. Under field operations, Dynamics Diff. Kinematics Kinematics
dust, temperature change, humidity variation, and unexpected
collision lead to noisy and drifting sensor data. Therefore, Therefore, this term is build up from dynamics, differential
Hwangbo et.al. have developed a method based on a HMM kinematics, and kinematics contributions. The differential
(Hidden Markov Model) [4] to estimate the contact state kinematics term P (q̇|Xt , q) highly depends on the specific
between foot and ground. The model used can be sepa- trajectory of the robot arm. Moreover, the kinematics term
rated into measurement model and transition model. The P (q|Xt ) depends on the trajectory and the prior information
measurement model focuses on the kinematics, differential of the manipulation environment. Therefore, to generalize
kinematics, and dynamics of the robot, and the transition the model for robotic arms manipulated under arbitrary
model focuses on the transition from the former state to the trajectory, we only consider the dynamics term in our work
current one. Their method has been tested on the quadrupedal and assume the other terms are uniform:
ANYmal [6] and achieved reliable results. Another method P (q, q̇, q̈|Xt ) ∝ P (q̈|Xt , q, q̇). (4)
is based on logistic classifier [5]. The classifier tries to
learn the Ground Reaction Force (GRF) threshold, which has This results in P (Xt ) and P (q̈|Xt , q, q̇) being the only terms
the highest probability to minimize the base velocity error. that are required for computing the probability of contact.
Moreover, the algorithm also merges individual legs’ velocity The computation approach is illustrated in the following.
1) P (q̈|Xt = N, q, q̇): In the case the arm moves in the Through linearity, we can get the distribution of estimated
air without contact, the dynamics can be summarized as acceleration given there is a contact:
M(q, q̇)q̈ + C(q, q̇)q̇ + g(q) = τj , (5) ˆq̈ ∼ N (M−1 (τj − Cq̇ − g − Jc Fe ),
C
(15)
M−1 JTc σFe J−1
c (M
−1 T
) + M−1 στ (M−1 )T + σq̈ ) .
where M(q, q̇) represents the mass matrix, C(q, q̇)q̇ repre-
sents the nonlinear term, g(q) represents the gravity term and Therefore, the conditional probability can be computed as
τj represents the joint torques. When noise is considered in
fˆq̈ (q̈) := P (q̈|Xt = C, q, q̇). (16)
the model, the estimated acceleration given no contact ˆq̈N C

can be computed as 3) P (Xt = N ) and P (Xt = C): The prior probability of


ˆ contact or not contact is highly dependent on the predefined
q̈N = M−1 (q, q̇)(τj − C(q, q̇)q̇ − g(q) + ωd ) + ωq̈ , (6)
trajectory of the robot. In our case, we assume no prior
where ωq̈ is the estimation noise in acceleration and ωd is knowledge about the trajectory and set the probability for
the combined noise from other sources. They are assumed to contact as well as not contact to 0.5.
be independent variables. To simplify the estimation model, B. Contact Force Estimation through Reference Signal
we assume ωq̈ and ωd are Gaussian distributed with zero Tracking
mean:
Computation of external wrench by Equation (10) directly
ωq̈ ∼ N (0, σq̈ ) , ωd ∼ N (0, στ ) , (7)
introduces noise due to double differentiation of the joint
where the variances are diagonal matrices of appropriate positions. This is why De Luca et.al’s work [1] uses the in-
dimension. Therefore, the estimated acceleration given no tegral over a generalized momentum to avoid the use of joint
contact ˆ
q̈N is also Gaussian distributed due to linearity accelerations. Their residual approach can also be interpreted
as signal tracking with only integral term, thereby inherently
ˆ
q̈N ∼ N (M−1 (q, q̇)(τj − C(q, q̇)q̇ − g(q)),
(8) introducing a phase shift. We extend this formulation by
M−1 (q, q̇)στ (M−1 (q, q̇))T + σq̈ ) . also adding a proportional term to have a tunable trade-off
between noise and delay. The approach can be summarized
We can then label the probability density function as
as Z
fˆq̈ (q̈) := P (q̈|Xt = N, q, q̇),
N
(9) τ̃ = Kp (τe − τ̃ ) + KI (τe − τ̃ )dt, (17)
where fˆq̈ is the probability density function, and the gener- where τe is the resulting torque at joint level from an external
N
alized acceleration q̈ can be computed from the generalized wrench Fe . τe can be represented by the system dynamics.
velocity q̇ by numerical differentiation. As shown in Equation (10),
2) P (q̈|Xt = C, q, q̇): In the case the arm is in contact,
the dynamics can be represented as τe = JTc Fe = τj − M(q, q̇)q̈ − C(q, q̇)q̇ − g(q). (18)

M(q, q̇)q̈ + C(q, q̇)q̇ + g(q) + JTc Fe = τj . (10) The integration term of Equation (17) acts as an first order
low pass filter, which assures that, τ̃ is less noisy than τe .
where Jc represents the contact Jacobian matrix and Fe However, the term also introduces time delay. The propor-
represents external wrench to the end-effector. The contact tional term can be used to eliminate time delay partly, but
wrench can be written as a function of the joint torque there is always a trade-off between time delay and precision.
Fe = (Jc M−1 JTc )−1 [J˙c q̇ + Jc M−1 (τj − Cq̇ − g)]. (11) C. Hybrid Approach of Contact Detection and Force Esti-
mation
This equation originates from contact constraints
While the estimation by reference signal tracking provides
r˙c = Jc q̇ = 0, (12) a good approximation of the interaction force in a contact sit-
uation, it is prone to model errors and introduces an inherent
where r˙c represents the velocity of end-effector in base time delay. Therefore it is not suitable to decide if a contact
frame. We assume that the constraint (12) is still satisfied occurs during a dynamic motion. The probabilistic contact
when the acceleration of the end-effector is small. The detection on the other hand can identify a contact situation
estimated acceleration given contact ˆ
q̈c can then be computed but provides no information on the corresponding force.
similar as Equation (6): Therefore it is reasonable to unify it with the estimation
ˆq̈ = M−1 (τj − Cq̇ − g + ωd − JT Fe + JT ωf ) + ωq̈ , (13) approach. This hybrid method results in fast detection and
C c c
accurate estimation of contact forces.
where the additional noise ωf is from the contact force A possible hybrid approach we propose for the task of
estimation, which is assumed to be independent from the desired end-effector force tracking is summarized in the
other two noise sources, and Fe can be substituted through pseudo code of Algorithm 1, which can be used for sensor-
Equation (11). ωf is also Gaussian distributed with zero less force tracking. It provides a set of rules on how to switch
mean: between pose and force control modes. pc is the probability
ωf ∼ N (0, σFe ) . (14) of contact, which is a scalar value and cp represents its
TABLE I
N OISE C HARACTERISTICS FOR I NDIVIDUAL J OINTS ( STARTS FROM
SHOULDER )

noise standard deviation


ωq̈ 25 rad/s2 , 25 rad/s2 , 25 rad/s2 , 25 rad/s2 , 25 rad/s2 , 25 rad/s2
ωd 0.15 Nm, 0.3 Nm, 0.3 Nm, 0.3 Nm, 0.3 Nm, 0.3 Nm
ωf 20 Nm, 30 Nm, 30 Nm, 30 N, 30 N, 30 N

TABLE II
C OMPARISON BETWEEN H YBRID A LGORITHM AND F ORCE S ENSOR
R EADING
Factors Hybrid Algorithm
Accuracy(deviation) ±10 %
Precision(noise) ±1.0
Fig. 2. Labeled joint configuration of ANYpulator. Time Response(delay) 0.1 s

threshold value. Fc represents the contact force at the end A. Experiment Setup
effector from the reference signal tracking, and cy is its The test platform used is ANYpulator, a 6 DoF lightweight
threshold value. Specific threshold values can be selected robot arm developed by the Robotic Systems Lab at ETH
for different application scenarios. Zurich, which is an extension of the work of [7]. The
When the robotic arm is manipulated under pose tracking arm is equipped with six series elastic actuators (SEA).
mode, as soon as the contact is detected by conditional The joint configuration of the arm is shown in Figure 2.
probability, the controller switches to force tracking mode. Due to the special mechanical design of the SEA, the joint
This mode remains as long as the criteria of contact satisfied torque can be acquired through measurement of the internal
by detection, or estimation. However, when both criteria are spring deflection. A 6 DoF force sensor, which can measure
not satisfied, then the control mode switches back to pose the 6 DoF wrench applied, is installed at the end of the
tracking mode. The relevant experiment is shown in Section link connected to the last actuator, providing benchmark
IV-C. measurements. It is important to note that ‘sensorless’ here
means the external wrench applied to the end-effector is
Algorithm 1 Pseudo code of the hybrid algorithm for force only measured for benchmarking, but not used for contact
tracking application detection or force estimation. The system overview is shown
1: update pc and Fc in Figure 1. The noise parameters at joint level used for the
2: if control mode == pose control then probabilistic distributions during contact detection are shown
3: if pc > cp then in Table I.
4: control mode ← f orce control
B. Precision, Accuracy and Time Response Analysis
5: τj ← f orce tracking controller
6: end if In this section, the precision, accuracy and time response
7: else of the proposed algorithm are analyzed on the platform. The
8: if pc < cp and |Fc | < cy then analysis is divided into two parts. Firstly, we compare esti-
9: control mode ← pose control mation by time-varying signal tracking with the benchmark
10: τj ← pose tracking controller measurements provided by the force/toque sensor. Secondly,
11: end if we focus on the time response issues. An overview of the
12: end if comparison is shown in Table II.
13: set τj 1) Precision and Accuracy: During this experiment,
forces are applied to the end-effector by hand. The compar-
ison between time-varying signal tracking and force sensor
readings is shown in Figure 3. It can be observed that the
IV. E XPERIMENTS AND R ESULTS estimation provides relatively accurate approximation of the
wrench, with reasonable noise amplitude of around ±1 (note
In this section we present two experiments performed the different scales for the force plots). The obvious noise
on a robot arm to verify the effectiveness of the proposed in the plot of the momentum around the y−axis originates
algorithms. The first one focuses on the precision, accuracy from the increased compliance in this direction due to the
and time delay issues of the hybrid algorithm. The second arm configuration.
aims at the force tracking application without the use of any 2) Time Response: In terms of time response, the analysis
extra force/torque sensors. All experiments are based on the is based on the hand guiding experiments, whose details are
fact that the external force is only applied to the end-effector. explained in Section IV-C. From Figure 5, we can see that
Fx
Contact Force [N]

0
-20
-40
-60

5 10 15 20 25
Time [s]
Fy
20
Contact Force [N]

Sensor Measurement
Estimation
0

-20
5 10 15 20 25
Time [s]
Fz
Contact Force [N]

10

-10

5 10 15 20 25
Time [s]

Mx
Contact Torque [Nm]

20
20
Conditional Probability
0 Signal Tracking Estimation
Desired Force
Force Sensor Measurement
-20
5 10 15 20 25 15
Time [s]
My
20
Torque [Nm]

10
0

Fx [N]
-20
5 10 15 20 25
5
Time [s]
Mz
20 Sensor Measurement
Torque [Nm]

Estimation
0 0

-20
5 10 15 20 25
Time [s] -5
10 12 14 16 18 20 22 24
Time [s]

Fig. 3. Comparison between sensor readings and force estimation by


reference signal tracking. All six motors of ANYpulator are in ‘freeze’ Fig. 4. Force tracking test on ANYpulator during hand guiding. Upper:
mode for this experiment. Four periods of the experiment, (1) Motion control (pose control mode); (2)
Contact detection; (3) Force control to track the desired end-effector force
(force control mode); (4) Contact released and keeping the current pose
(pose control mode). Lower: The desired end-effector force applying to the
the force sensor responds faster than the estimation through hand is set to 15 N. Observed oscillation of measurements and estimation
signal tracking and the probabilistic contact detection. How- originate from hand guiding. The variance of estimation at start originates
from the manipulator’s motion before contact.
ever, the estimated value for the contact force varies with the
manipulator’s motion even under no contact, due to model
inaccuracies (see from 10 s to 14 s in Figure 4). Therefore,
the threshold value for contact detection has to be set larger
than 5 N. The higher threshold value we set, the slower the 20
Conditional Probability
Signal Tracking Estimation

contact detection by signal tracking becomes compared to Desired Force


Threshold of Estimation
Force Sensor Measurement
the conditional probability response. In this manner, contact 15

detection by conditional probability responds faster than


estimation by signal tracking. 10

To conclude, the hybrid algorithm can detect contact with


Fx [N]

a response delay of approx. 0.1 s compared to force sensor 5

readings. Using the redundancy of the probabilistic contact


detection improves the response latency due to thresholding 0

of the contact force estimation through signal tracking. In


the next section, its application on compliant force tracking -5
11 12 13 14 15 16 17
will be demonstrated. Time [s]

C. Force Tracking
The force tracking task is an application of the hybrid Fig. 5. Details of force tracking test on ANYpulator during hand guiding.
When the threshold value of estimation is considered, contact detection by
algorithm. Traditionally, the task is demonstrated by force conditional probability is actually faster.
control based on force sensor feedback. In this work, the task
the end-effector against fixed rigid objects. For this purpose,
14 the arm moves against a wall and applies a constant force of
Conditional Probability
Signal Tracking Estimation
Desired Force 6 N against it. The desired force that the end-effector should
12 Force Sensor Measurement
apply to the wall is transformed back to the corresponding
10
joints as desired joint torque (see Figure 6). The pink lines
8 show the torque that is needed at joint level to apply the
desired contact force and the red lines represent the torque
Fx [N]

from force estimation. The task is achieved successfully with


4
minor errors from modeling error, offset of force sensors
2 or misalignment between force direction and force sensor’s
0
main axis.
-2
10 15 20 25 30 35
V. C ONCLUSION
Time [s]
This work introduced a hybrid algorithm of sensorless
contact detection and contact force estimation for robotic
arms with general manipulation tasks. The algorithm is
0.2
SH ROT
Conditional Probability
1.5
SH FLE
0.5
EL FLE
composed of two parts, fast contact detection by conditional
0
Estimation Torque
Desired Torque
1 0
probability and contact estimation by signal tracking. The
SH ROT [Nm]

SH FLE [Nm]

accuracy, precision, and time response of the algorithm have


EL FLE [Nm]

-0.2

-0.4 0.5 -0.5

-0.6 been analyzed and evaluated with benchmark measurements


0 -1
-0.8
provided by a force/torque sensor. With the reliable algo-
-1 -0.5 -1.5
15 20
time [s]
25 30 35 15 20 25
time [s]
30 35 15 20 25
time [s]
30 35
rithm proposed, the compliant force tracking task has been
0.2
WR FLE
0.1
WR DEV
1
WR PRO achieved successfully on ANYpulator without the use of any
0
0
0.5
extra force/torque sensors.
WR PRO [Nm]
WR DEV [Nm]
WR FLE [Nm]

-0.1

-0.2 -0.2 0 ACKNOWLEDGEMENT


-0.3
-0.4
-0.4
-0.5
The authors would like to thank Hendrik Kolvenbach and
-0.6
15 20 25 30 35
-0.5
15 20 25 30 35
-1
15 20 25 30 35 Klajd Lika of RSL for their work on force sensors.
time [s] time [s] time [s]

R EFERENCES
[1] A. De Luca, A. Albu-Schaffer, S. Haddadin, and G. Hirzinger, “Colli-
Fig. 6. Force tracking test on ANYpulator when guided to push the fixed sion detection and safe reaction with the DLR-III lightweight manipula-
rigid objects with desired force of 6 N. The upper plot shows the force at tor arm,” in International Conference on Intelligent Robots and Systems
the end-effector level, and the lower plot shows the corresponding torque (IROS). IEEE, 2006, pp. 1623–1630.
at joint level. [2] S.-D. Lee, K.-H. Ahn, and J.-B. Song, “Torque control based sensorless
hand guiding for direct robot teaching,” in International Conference on
Intelligent Robots and Systems (IROS). IEEE, 2016, pp. 745–750.
is completed without using the force sensors. For the detailed [3] Y. Tian, Z. Chen, T. Jia, A. Wang, and L. Li, “Sensorless collision
detection and contact force estimation for collaborative robots based
control algorithm, please refer to Algorithm 1. Basically, if on torque observer,” in International Conference on Robotics and
during motion (pose tracking mode, as (1) of Figure (4)), Biomimetics (ROBIO). IEEE, 2016, pp. 946–951.
an external force is detected at the end-effector (2) , the [4] J. Hwangbo, C. D. Bellicoso, P. Fankhauser, and M. Huttery, “Proba-
bilistic foot contact estimation by fusing information from dynamics
controller switches to the force tracking mode and tracks the and differential/forward kinematics,” in International Conference on
desired end-effector force. The end-effector can apply the Intelligent Robots and Systems (IROS). IEEE, 2016, pp. 3872–3878.
desired force either to a fixed rigid object or to a moving [5] M. Camurri, M. Fallon, S. Bazeille, A. Radulescu, V. Barasuol, D. G.
Caldwell, and C. Semini, “Probabilistic contact estimation and impact
object, like a guiding hand, as (3). As soon as the external detection for state estimation of quadruped robots,” IEEE Robotics and
guiding is released, the controller switches back to the pose Automation Letters, vol. 2, no. 2, pp. 1023–1030, 2017.
control mode and tracks the current end-effector pose (4). [6] M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V. Tsounis,
J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch et al., “Anymal-
The result of force tracking during hand guiding is shown a highly mobile and dynamic quadrupedal robot,” in International
in Figure 4. Along the guiding, the controller tracks the Conference on Intelligent Robots and Systems (IROS). IEEE, 2016,
constant desired force of 15 N with reasonably fast response. pp. 38–44.
[7] K. Bodie, C. D. Bellicoso, and M. Hutter, “Anypulator: Design and
When the hand guiding is released, it switches back to control of a safe robotic arm,” in International Conference on Intelligent
task space control of the end effector. What should be Robots and Systems (IROS). IEEE, 2016, pp. 1119–1125.
mentioned is that the conditional probability does always
provide ‘contact’ during the motion. Additionally, high signal
tracking estimation indicates external wrench. This motivates
us to combine contact detection by conditional probability
with estimation by reference signal tracking together as
release criteria for the hybrid algorithm.
Furthermore, we tested the force tracking when pushing

You might also like