scispace - formally typeset
Search or ask a question
Proceedings ArticleDOI

Collision Detection and Safe Reaction with the DLR-III Lightweight Manipulator Arm

TL;DR: An efficient collision detection method that uses only proprioceptive robot sensors and provides also directional information for a safe robot reaction after collision is presented.
Abstract: A robot manipulator sharing its workspace with humans should be able to quickly detect collisions and safely react for limiting injuries due to physical contacts. In the absence of external sensing, relative motions between robot and human are not predictable and unexpected collisions may occur at any location along the robot arm. Based on physical quantities such as total energy and generalized momentum of the robot manipulator, we present an efficient collision detection method that uses only proprioceptive robot sensors and provides also directional information for a safe robot reaction after collision. The approach is first developed for rigid robot arms and then extended to the case of robots with elastic joints, proposing different reaction strategies. Experimental results on collisions with the DLR-III lightweight manipulator are reported.

Summary (3 min read)

Introduction

  • Safety issues are of primary concern when robot manipulators are supposed to operate in an unstructured environment, sharing the workspace with a human user and allowing a close physical cooperation [1], [2].
  • Along a similar but simpler line, robot manufacturers are providing hw/sw facilities for prescribing safe Cartesian areas that should not be accessed by the robot in any operative condition.
  • Tuning of collision detection thresholds in these schemes is difficult because of the highly varying dynamic characteristics of the commanded torques.
  • A detection scheme that works under similar conditions and avoids the above drawbacks has been recently proposed in [15].
  • Collisions are viewed as faulty behaviors of the robot actuating system, while the design of a detector takes advantage of the decoupling property of robot generalized momentum [16], [17].

II. PRELIMINARIES

  • The authors first consider robot manipulators as open kinematic chains of rigid bodies, having N rigid joints.
  • The dynamic model is M(q)q̈ + C(q, q̇)q̇ + g(q) = τ tot, (1) where M(q) is the symmetric, positive definite inertia matrix, the Coriolis and centrifugal terms are factorized using the matrix C(q, q̇) of Christoffel symbols, and g(q) is the gravity vector.
  • In the absence of friction and other external torques acting from the environment, τ tot is just the motor torque τ .
  • The motor inertia matrix B and the joint stiffness matrix K are diagonal and positive definite, while D ≥ 0 is the diagonal joint viscosity matrix.
  • The authors note that the analysis presented in this paper applies, with minor modifications, also to the more complete model of robots with elastic joints considered, e.g., in [20].

III. DETECTION AND REACTION WITH RIGID ROBOTS

  • During normal operation, the robot arm may collide with a standing or moving person/obstacle in its workspace.
  • Accordingly, the Cartesian collision forces and moments are denoted by F K = [ fK mK ] ∈ R6.

A. Collision detection

  • Note that σ can be computed using the measured joint position q and velocity q̇ (possibly obtained through numerical differentiation) and the commanded motor torque τ .
  • Not all possible collision situations are detected by this scheme.
  • With the robot at rest (q̇ = 0), the instantaneous value of τK does not affect σ, whereas this will happen only when the robot starts moving.
  • On the other hand, when the robot is in motion, collision can be detected provided that the Cartesian collision force produces motion at the contact.
  • In fact, no possible robot motion would be able to reduce the force loading in this case.

B. Collision identification

  • The previous scheme does not provide any directional information on the Cartesian collision force, nor is able to identify which robot link has collided.
  • In view of the structure (15), the authors call r a collision identification signal, or simply a residual bearing this term from the fault detection literature.
  • More in general, the sensitivity to F K of each of the affected residuals (proximal to the robot base) will vary with the arm configuration (see also [15]).
  • Thanks to the properties of the generalized momentum, this dynamic analysis can be carried out based only on the static transformation matrix JTK(q) from Cartesian forces to joint torques.
  • In fact, the residual dynamics in eq. (14) is unaffected by robot velocity and acceleration.

C. Reaction strategy

  • Enforcing such a zero-gravity condition is useful for guaranteeing a safe behavior [1], as robot motion will not be biased along the gravity direction.
  • During normal operation (pre-impact phase), it is convenient to apply a control law that provides accurate trajectory tracking in free motion while displaying passive properties (springdamper type) in response to unexpected collisions.
  • The simplest reaction strategy to a collision would be to stop the robot by using its brakes [6].
  • Instead, the directional information embedded in the residual vector r can be used in the post-impact phase, by switching control to a more friendly behavior.
  • Equation (18) is clearly an active control scheme, as additional energy is feed into the system after the collision has been detected.

D. Energy dissipation

  • The robot operation states and their transition conditions are shown in Fig 1.
  • Note that as long as the collision flag is up, any further collision will keep the robot in the reflex reactive state.

IV. EXTENSION TO ROBOTS WITH JOINT ELASTICITY

  • The extension of the collision detection and identification schemes developed in Sect. III for the rigid case can be made in different ways.
  • In view of the application of their methods to the DLR-III lightweight manipulator, the authors take advantage of the specific sensing devices available on board of this robot.
  • In particular, every joint is equipped with a high-resolution incremental position sensor on the motor side and an integrated joint torque sensor, so that θ and τ J are directly available.
  • More specifically, substituting τ with τ J + DK−1τ̇ J in the expressions (11) and (13), the authors obtain similar collision detection schemes and identification properties.
  • This leads again to the linear and decoupled residual dynamics ṙEJ = −KIrEJ + KIτK .

A. DLR-III controller

  • The authors recall first the form of the robot control law used for general tasks with the DLR-III arm, in which reaction strategies to collisions have been inserted as additional control modalities.
  • The authors note that the passivity properties are preserved provided that the following condition on the introduced damping matrices is satisfied [18]: D ≥ 1 4 (Ds −D)T Dθ (Ds −D).
  • Equation (25) represents a full-state feedback with respect to suitable, possibly time-varying, reference values.
  • This fact will be used for implementing different strategies of robot reaction to collisions in a single framework.
  • Realizing a zero-gravity condition for an elastic joint robot, with a choice equivalent to eq. (16) of the rigid case, is not as immediate.

B. Reaction strategies

  • The residual rEJ in eq. (22) is used for detecting a collision and identifying a safe direction for the reactive motion of the robot.
  • As baseline behaviors in the performed collision experiments, the authors have taken the case of no reaction at all (Strategy 0) and of immediate stop of the trajectory generation with simultaneous high-gain position control (Strategy 1).
  • With reference to the general control law (25), three reactive strategies have been considered.
  • This strategy leaves the robot floating in space in response to the collision force, while motion is damped at the motor side.
  • This strategy is the closest to eq. (18) of the rigid case.

V. EXPERIMENTS WITH THE DLR-III MANIPULATOR

  • The authors have performed several tests on collision detection with the DLR-III lightweight manipulator and using the robot reaction strategies defined in Section IV-B.
  • The authors report here numerical results obtained on repeated collisions with a balloon (see Figs. 2–6) and qualitative results for collisions with humans (see Fig. 7).
  • Fig. 2. Collision with a balloon (motion at 100◦/s) and robot reaction Collisions occurred between the hand or the outstretched arm of different test persons and different locations between the 4-th and 6- th link of the robot, with linear speeds at the contact up to approximately 1.5 m/s.
  • When the time interval of contact is relatively long, the detection and reaction capabilities of the robot are enhanced.

VI. CONCLUSIONS

  • The authors have presented a complete approach, from detection to reaction, for handling human-robot collisions without the need of external sensing.
  • Collision detection and identification signals can be efficiently generated resorting to energy arguments or based on the robot generalized momentum and by using only proprioceptive measurements.
  • The robot retracts itself safely and rapidly away from the collision area, using the local directional information collected during the impact.
  • On-going work is concerned with acceleration-driven collision detection and the reduction of control communication delays in their robotic set-up, as well as with a more accurate evaluation of several severity indices of the impacts and of the beneficial inclusion of compliant coverages.
  • Furthermore, robot redundancy will be exploited in order to devise reaction strategies that try to complete a given Cartesian motion task, despite of the detected collision.

Did you find this useful? Give us your feedback

Content maybe subject to copyright    Report

Collision Detection and Safe Reaction
with the DLR-III Lightweight Manipulator Arm
Alessandro De Luca
Dipartimento di Informatica e Sistemistica
Universit
`
a di Roma “La Sapienza”
Via Eudossiana 18, 00184 Roma, Italy
deluca@dis.uniroma1.it
Alin Albu-Sch
¨
affer Sami Haddadin Gerd Hirzinger
DLR - German Aerospace Center
Institute of Robotics and Mechatronics
P.O. Box 1116, D-82230 Wessling, Germany
{alin.albu-schaeffer, sami.haddadin, gerd.hirzinger}@dlr.de
Abstract A robot manipulator sharing its workspace with
humans should be able to quickly detect collisions and safely
react for limiting injuries due to physical contacts. In the absence
of external sensing, relative motions between robot and human
are not predictable and unexpected collisions may occur at
any location along the robot arm. Based on physical quantities
such as total energy and generalized momentum of the robot
manipulator, we present an efficient collision detection method
that uses only proprioceptive robot sensors and provides also
directional information for a safe robot reaction after collision.
The approach is first developed for rigid robot arms and then
extended to the case of robots with elastic joints, proposing
different reaction strategies. Experimental results on collisions
with the DLR-III lightweight manipulator are reported.
I. INTRODUCTION
Safety issues are of primary concern when robot manipula-
tors are supposed to operate in an unstructured environment,
sharing the workspace with a human user and allowing a close
physical cooperation [1], [2]. Accidental collisions that may
harm humans should be avoided by anticipating dangerous
situations, while the effects of actual collisions should be
mitigated by having the robot react promptly so as to recover
a safe operative condition.
Current research on physical human-robot interaction deals
with different aspects involved in robot dependability: me-
chanical design, aimed at reducing manipulator inertia and
weight; introduction of compliant components, for reducing
the severity of impacts; additional use of external sensors,
so as to allow a fast detection of human-robot proximity;
motion planning and control strategies, for minimizing the
risks associated to collisions. Each of these aspects is relevant
in one or more of the elementary phases in which a physical
human-robot interaction can be divided.
In the pre-impact phase, collision avoidance is the primary
goal and requires (at least, local) knowledge of the current
environment geometry and computationally intensive motion
planning techniques. Along a similar but simpler line, robot
manufacturers are providing hw/sw facilities for prescribing
safe Cartesian areas that should not be accessed by the robot
in any operative condition. Anticipating incipient collisions or
recognizing them in real-time is typically based on the use
of additional external sensors, such as sensitive skins [3], on-
board vision [4], strain gages, force load cells, and so on.
When a collision occurs, the resulting contact forces during
the impact phase may be alleviated by pursuing a lightweight
robot design [5], by adding soft visco-elastic covering to
the links [6], or by introducing compliance in the driving
system so as to mechanically decouple the heavy motor inertia
from the link inertia [7]–[9]. Light but stiff link materials
can be combined with harmonic drives, introducing thus joint
elasticity, as in the DLR-III robot manipulator. Also, variable
stiffness actuation can be used, stiffening the joints during
low-velocity transients and relaxing them at high-velocity
regimes [10], in such a way that tracking performance during
free motion is not compromised by the introduced compliance.
In the post-impact phase, the first task is to detect the
collision occurrence, which may have happened at any location
along the robot arm. The controller should then switch to an
appropriate reaction strategy, the most simple one being to
stop the robot. It is obviously more cost effective to be able
to detect a collision without the need of additional sensors. A
rather intuitive scheme is to compare the commanded torque
(or, the current in an electrical drive) with the nominal model-
based command (i.e., the torque expected in the absence
of collision) and looking for fast transients due to possible
collision [6], [11], [12]. This approach has been refined by
including adaptive compliance control [13], [14]. However,
tuning of collision detection thresholds in these schemes is
difficult because of the highly varying dynamic characteristics
of the commanded torques. Moreover, their common draw-
back (even when robot dynamics is perfectly known) is that
the inverse dynamics computation for torque comparison is
based on acceleration estimates that introduce noise (due to
numerical differentiation of velocity or position data) and/or
an intrinsic delay in a digital implementation.
A detection scheme that works under similar conditions
and avoids the above drawbacks has been recently proposed
in [15]. Collisions are viewed as faulty behaviors of the
robot actuating system, while the design of a detector takes
advantage of the decoupling property of robot generalized
momentum [16], [17]. Moreover, this detection scheme is
particularly convenient when it is necessary to switch control
strategy, being independent from the control method used to
generate the commanded motor torques.
In this paper, we build upon the original idea of [15]
1-4244-0259-X/06/$20.00 ©2006 IEEE
1623
Proceedings of the 2006 IEEE/RSJ
International Conference on Intelligent Robots and Systems
October 9 - 15, 2006, Beijing, China

and present a complete treatment of the post-impact phase,
from collision detection and identification to robot reaction
strategies. In particular, the directional information on contact
forces provided by the identification scheme is used to safely
drive the robot away from the human. Section II recalls
the physical properties of the robot total energy and of its
generalized momentum which are relevant for our problem.
New detection schemes and an analysis of the detectability
of collision forces are presented in Section III for rigid
manipulators. The extension to the case of robots with elastic
joints, with alternative sets of reaction strategies, is discussed
in Section IV. Finally, Section V reports on the experimental
results obtained with the DLR-III lightweight manipulator.
II. PRELIMINARIES
We first consider robot manipulators as open kinematic
chains of rigid bodies, having N (rotational) rigid joints. The
generalized coordinates q R
N
can be associated to the
position of the links. The dynamic model is
M (q)
¨
q + C(q,
˙
q)
˙
q + g(q) = τ
tot
, (1)
where M (q) is the symmetric, positive definite inertia matrix,
the Coriolis and centrifugal terms are factorized using the
matrix C(q,
˙
q) of Christoffel symbols, and g(q) is the gravity
vector. In the right-hand side of (1), τ
tot
contains all active
generalized torques performing work on q and all dissipative
torques. In the absence of friction and other external torques
acting from the environment, τ
tot
is just the motor torque
τ . From the skew-symmetry of matrix
˙
M (q) 2C(q,
˙
q) it
follows that
˙
M (q) = C(q,
˙
q) + C
T
(q,
˙
q). (2)
The total energy of the robot is the sum of its kinetic energy
and potential energy due to gravity:
E = T + U =
1
2
˙
q
T
M (q)
˙
q + U
g
(q), (3)
with g(q) = (U
g
(q)/∂q)
T
. From (1) and (2), it is
˙
E =
˙
q
T
τ
tot
, (4)
which represents the energy balance in the system.
The generalized momentum of the robot is defined as
p = M (q)
˙
q. (5)
Using again (1) and (2), its time evolution is given by
˙
p = τ
tot
+ C
T
(q,
˙
q)
˙
q g(q). (6)
The i-th component of
˙
p can also be written as
˙p
i
= τ
tot,i
1
2
˙
q
T
M (q)
q
i
˙
q g
i
(q),
for i = 1, . . . , N. The dynamics of the generalized momentum
is thus decoupled component-wise with respect to the torques
acting on the right-hand side of eq. (1).
In the second part of the paper, we will extend our
results to the case of robots with rigid links but elastic
joints/transmissions. For the dynamic modeling of elastic joint
robots, a doubling of generalized coordinates is needed in a
Lagrangian formulation. Let θ R
N
be the motor positions
(as reflected through the gearboxes) and q R
N
be the
link positions. Elasticity of the motion transmission elements
is modeled by linear springs introduced at each joint. We
consider also joint viscosity effects as in [18]. With the
standard assumptions in [19], the dynamic model becomes
M (q)
¨
q + C(q,
˙
q)
˙
q + g(q) = τ
tot,J
B
¨
θ + DK
1
˙
τ
J
+ τ
J
= τ ,
(7)
where
τ
J
= K(θ q) (8)
is the elastic force transmitted through the joints. The motor
inertia matrix B and the joint stiffness matrix K are diagonal
and positive definite, while D 0 is the diagonal joint
viscosity matrix. In the right-hand side of the first equation
in (7), τ
tot,J
contains the torques performing work on q,
i.e., those transmitted through the visco-elastic joints and the
external torques acting from the environment. When the latter
are absent, it is τ
tot,J
= τ
J
+ DK
1
˙
τ
J
= K(θ q) +
D(
˙
θ
˙
q). We note that the analysis presented in this paper
applies, with minor modifications, also to the more complete
model of robots with elastic joints considered, e.g., in [20].
III. DETECTION AND REACTION WITH RIGID ROBOTS
During normal operation, the robot arm may collide with
a standing or moving person/obstacle in its workspace. For
simplicity, we assume that there is at most a single link
involved in the collision. Let
V
K
=
v
K
ω
K
=
J
K,lin
(q)
J
K,ang
(q)
˙
q = J
K
(q)
˙
q R
6
be the linear velocity at the contact point and the angular
velocity of the associated robot link. The quantity V
K
, and
in particular the (geometric) Jacobian J
K
(q), is unknown
in advance. Accordingly, the Cartesian collision forces and
moments are denoted by
F
K
=
f
K
m
K
R
6
.
When a collision occurs, the robot dynamics (1) becomes
M (q)
¨
q + C(q,
˙
q)
˙
q + g(q) = τ + τ
K
, (9)
where the joint torque τ
K
associated to the Cartesian collision
(generalized) force F
K
is given by
τ
K
= J
T
K
(q)F
K
. (10)
A. Collision detection
Define the scalar quantity
σ(t) = k
D
E(t)
Z
t
0
˙
q
T
τ + σ
ds E(0)
, (11)
with σ(0) = 0, k
D
> 0, and where E(t) is the total robot
energy at time t 0, as defined in (3). Note that σ can be
1624

computed using the measured joint position q and velocity
˙
q
(possibly obtained through numerical differentiation) and the
commanded motor torque τ . The latter may be the result of
any type of control action. No acceleration measurement is
needed.
Using eqs. (4) and (9), the resulting dynamics of σ is
˙σ = k
D
σ + k
D
˙
q
T
τ
K
, (12)
i.e., that of a first-order stable linear filter driven by the work
performed by the joint torques due to collision. During free
motion, σ = 0 up to measurement noise and unmodeled
disturbances. In response to a generic collision, σ raises
exponentially with a time constant 1/k
D
and detection occurs
as soon as |σ| > σ
low
, a suitable threshold whose actual value
depends on the noise characteristics in the system. Dynamic
thresholding can be used for avoiding false detection due to
spurious spikes in noisy signals, as shown in [21]. When
contact is lost, σ rapidly returns to zero. Because of these
properties, we call σ a collision detection signal.
Not all possible collision situations are detected by this
scheme. With the robot at rest (
˙
q = 0), the instantaneous
value of τ
K
does not affect σ, whereas this will happen only
when the robot starts moving. As a consequence, with the
robot at rest, true impulsive collision forces/torques cannot be
detected by this scheme, unless they are mechanically filtered
by the presence of a soft covering of the arm. On the other
hand, when the robot is in motion, collision can be detected
provided that the Cartesian collision force produces motion at
the contact. In fact, using eq. (10),
˙
q
T
τ
K
=
˙
q
T
J
T
K
(q)F
K
= V
T
K
F
K
= 0 V
K
F
K
.
As a simple instance, a lateral (horizontal) force due to a
human colliding against a 2R planar arm in motion in the
vertical plane will not be detected, being fully compensated
by the reaction forces of the manipulator structure. When
evaluated in terms of reactive motions that the robot may take
in response to this collision, such behavior of the detection
scheme is rather natural. In fact, no possible robot motion
would be able to reduce the force loading in this case. Suppose
now to add a vertical joint axis at the base (obtaining a 3R
elbow-type robot) and let the second and third links be in
motion in the vertical plane as before (i.e., with ˙q
1
= 0). The
same previous lateral force will be felt initially only at the first
joint (τ
K,1
6= 0), which is however at rest, so that
˙
q
T
τ
K
= 0
and thus ˙σ(0) = 0. Provided that the joint position controller
is soft enough, the first joint will start moving in response to
the collision before the contact force has been removed and
detection may then occur.
B. Collision identification
The previous scheme does not provide any directional infor-
mation on the Cartesian collision force, nor is able to identify
which robot link has collided. To this purpose, following [15],
we define the N-dimensional quantity
r(t)=K
I
p(t)
Z
t
0
τ + C
T
(q,
˙
q)
˙
q g(q) + r
ds p(0)
(13)
with r(0) = 0, a diagonal gain matrix K
I
> 0, and where
p(t) is the robot generalized momentum at time t 0, as
defined in (5). Vector r can be computed using the measured
(q,
˙
q) and the commanded motor torque τ . In particular, no
inversion of the inertia matrix is needed.
From eqs. (6) and (9), the dynamics of r is
˙
r = K
I
r + K
I
τ
K
, (14)
or, component-wise in the Laplace domain,
r
j
(s)
τ
K,j
(s)
=
K
I,j
s + K
I,j
, j = 1, . . . , N,
with the N decoupled transfer functions having unitary gains.
For (each component of) r, all the appealing properties
of the scalar detection signal σ hold as well. In particular,
collision will now be detected when krk > r
low
or, by working
component-wise, when there exists at least an index j, with
j = 1, . . . , N, for which |r
j
| > r
low,j
. In ideal conditions,
K
I
r τ
K
,
which means in practice that the gains should be taken as
large as possible. Moreover, r is sensitive to collisions even
at
˙
q = 0. When the contact occurs on the i-th link of the robot
kinematic chain, we have
r =
. . . 0 . . . 0
T
.
i + 1 . . . N
(15)
Assuming r τ
K
= J
T
K
(q)F
K
, this follows from the fact
that, for a collision on link i, the last N i columns of the Jaco-
bian J
K
(q) are identically zero. In view of the structure (15),
we call r a collision identification signal, or simply a residual
bearing this term from the fault detection literature. The first i
components of vector r will be generically different from zero,
at least for the time interval of contact, and will start decaying
exponentially toward zero as soon as contact is lost. The
residual r will be affected only by Cartesian collision forces
F
K
that perform virtual work on admissible robot motion,
i.e., those forces that do not belong to the kernel of J
T
K
(q).
More in general, the sensitivity to F
K
of each of the affected
residuals (proximal to the robot base) will vary with the arm
configuration (see also [15]). Thanks to the properties of the
generalized momentum, this dynamic analysis can be carried
out based only on the static transformation matrix J
T
K
(q) from
Cartesian forces to joint torques. In fact, the residual dynamics
in eq. (14) is unaffected by robot velocity and acceleration.
C. Reaction strategy
In general, we suppose that the robot is always gravity
compensated, i.e.,
τ = τ
0
+ g(q), (16)
1625

with τ
0
given by any motion control law. Enforcing such
a zero-gravity condition is useful for guaranteeing a safe
behavior [1], as robot motion will not be biased along the
gravity direction. When motor torques are provided by a
control law of the form (16), the computation of the detection
signal σ and of the residual r may be simplified. In (11), τ can
be replaced by τ
0
and the total energy E by the kinetic energy
T , while in (13) one could just use τ
0
in place of τ g(q).
These modifications will not alter the resulting dynamics (12)
(linear for σ) and (14) (linear and decoupled for r).
During normal operation (pre-impact phase), it is convenient
to apply a control law that provides accurate trajectory tracking
in free motion while displaying passive properties (spring-
damper type) in response to unexpected collisions. A typical
choice is a joint PD linear feedback law
τ
0
= τ
0
d
+ K
P
(q
d
q) K
D
˙
q, (17)
with (diagonal) K
P
> 0 and K
D
> 0, and with the nominal
feedforward term computed on the desired trajectory q
d
(t) as
τ
0
d
= M (q
d
)
¨
q
d
+ C(q
d
,
˙
q
d
)
˙
q
d
+ K
D
˙
q
d
.
The simplest reaction strategy to a collision would be to
stop the robot by using its brakes [6]. However, this would not
remove the arm from direct contact with a human, generating
an unpleasant feeling of permanent danger or even squeezing
the person in a narrow environment. Instead, the directional
information embedded in the residual vector r can be used
in the post-impact phase, by switching control to a more
friendly behavior. In response to a collision, the residual r
rapidly increases and reaches peak values that depend on the
severity of the impact. As soon as the detection threshold(s)
is exceeded, the control law (17) is switched to the following
simple reflex strategy:
τ
0
= K
R
r, (diagonal) K
R
> 0. (18)
The idea is to use the motor torques so as to (over)react to
the external collision force along the same resulting direction,
as seen at the level of joint torques. By combining eqs. (9),
(16), and (18), and assuming the limit case of r = τ
K
, after
collision detection the robot dynamics becomes
(I + K
R
)
1
(M (q)
¨
q + C(q,
˙
q)
˙
q) = τ
K
. (19)
The robot inertial terms, as seen by a collision torque, are
scaled by a factor larger than unity —a lighter pushable robot
is obtained. In practice, because of the filtering introduced
by eq. (14), this result will be obtained only partially. Equa-
tion (18) is clearly an active control scheme, as additional
energy is feed into the system after the collision has been
detected. The expected outcome, however, is that the robot
bounces back in a direction (implicitly defined by eq. (18)
and the robot arm dynamics) which is the most advantageous
for escaping contacts, at least locally. When the contact is lost,
and in the absence of further collisions, the residual will return
to zero.
Fig. 1. Robot operation states
D. Energy dissipation
In the presence of very low friction, it may be necessary to
limit the excursion of the robot reflex motion. In such cases,
the control law (18) is only kept until krk r
stop
, when a
phase of maximum dissipation of kinetic energy is executed
in order to rapidly stop the robot in a safe configuration. Let
the available motor torques at each joint be bounded by
|τ
i
| τ
max,i
, i = 1, . . . , N.
Part of this motor torque is spent for the gravity compensation
in eq. (16). By defining the configuration-dependent bounds
τ
0
m,i
(q) = (τ
max,i
+ g
i
(q)) < 0
τ
0
M,i
(q) = τ
max,i
g
i
(q) > 0,
the remaining part of the available torques should satisfy
τ
0
m,i
(q) τ
0
i
τ
0
M,i
(q), i = 1, . . . , N.
Since the time evolution of the kinetic energy is
˙
T =
˙
q
T
τ
0
,
the following control law locally realizes the largest decrease
of T :
τ
i
=
τ
0
m,i
(q) if ˙q
i
ε
i
τ
0
m,i
(q) ˙q
i
i
if ε
i
> ˙q
i
0
τ
0
M,i
(q) ˙q
i
i
if ε
i
< ˙q
i
0
τ
0
M,i
(q) if ˙q
i
ε
i
+ g
i
(q),
with i = 1, . . . , N. For each velocity ˙q
i
, the insertion of a
small ultimate region of amplitude 2ε
i
> 0 allows a trade-
off between the almost minimum-time solution and a smooth
reaching of the final condition
˙
q = 0.
The robot operation states and their transition conditions are
shown in Fig 1. Note that as long as the collision flag is up,
any further collision will keep the robot in the reflex reactive
state. In other terms, the generator of the collision residual r
in eq. (13) is permanently active.
1626

IV. EXTENSION TO ROBOTS WITH JOINT ELASTICITY
For a serial robot manipulator with significant
joint/transmission elasticity, when a collision occurs the
dynamic model (7) becomes
M (q)
¨
q + C(q,
˙
q)
˙
q + g(q) = τ
J
+ DK
1
˙
τ
J
+ τ
K
(20)
B
¨
θ + DK
1
˙
τ
J
+ τ
J
= τ , (21)
with τ
K
= J
T
K
(q)F
K
as before.
The extension of the collision detection and identification
schemes developed in Sect. III for the rigid case can be
made in different ways. One possibility is to work with the
complete energy of the robot, by including the kinetic energy
of the motors and the potential energy associated to the joint
elasticity in the definition of a detection signal σ
EJ
, or by
considering also the generalized momentum associated to the
motors in the definition of a residual r
EJ
. Such schemes would
in principle require the measure of the whole state (θ,
˙
θ, q,
˙
q)
of the robot, together with the knowledge of the commanded
torque τ .
However, in view of the application of our methods to
the DLR-III lightweight manipulator, we take advantage of
the specific sensing devices available on board of this robot.
In particular, every joint is equipped with a high-resolution
incremental position sensor on the motor side and an integrated
joint torque sensor, so that θ and τ
J
are directly available.
From eq. (8), the link position is computed as q = θK
1
τ
J
.
Finally,
˙
q and
˙
τ
J
are obtained by numerical differentiation.
As a result, it is possible to consider only eq. (20) and
handle the measured joint torque τ
J
as the equivalent of
the commanded torque τ in the rigid robot model (9).
More specifically, substituting τ with τ
J
+ D K
1
˙
τ
J
in
the expressions (11) and (13), we obtain similar collision
detection schemes and identification properties. In particular,
the residual r
EJ
R
N
for the visco-elastic joint case is
defined as
r
EJ
(t) = K
I
p(t)
Z
t
0
τ
J
+ DK
1
˙
τ
J
+ C
T
(q,
˙
q)
˙
q g(q) + r
EJ
ds p(0)
i
,
(22)
with r
EJ
(0) = 0 and a (diagonal) gain matrix K
I
> 0. This
leads again to the linear and decoupled residual dynamics
˙
r
EJ
= K
I
r
EJ
+ K
I
τ
K
.
A. DLR-III controller
We recall first the form of the robot control law used
for general tasks with the DLR-III arm, in which reaction
strategies to collisions have been inserted as additional control
modalities. An interesting feature of this general controller,
used already for Cartesian compliance control [18], is the
possibility of shaping the motor inertia by a joint torque
feedback of the form
τ = BB
1
θ
u + (I BB
1
θ
)τ
J
+ (D BB
1
θ
D
s
)K
1
˙
τ
J
,
(23)
where u R
N
is the new torque command, B
θ
is the
desired (still diagonal) motor inertia, with 0 < B
θ
< B,
and D
s
> 0 is the desired joint torque damping (diagonal)
matrix. Substituting eq. (23) in (21) yields
B
θ
¨
θ + D
s
K
1
˙
τ
J
+ τ
J
= u.
As a result, the inertia of the robot motors appears to be
reduced —a convenient property for physical human-robot
interaction. In fact, the motor inertia B is a significant part
of the total inertia of the DLR-III manipulator: by reducing it
to B
θ
, the impact energy of the robot at a given velocity is
scaled accordingly. Moreover, the effect of friction will be also
reduced, allowing the human to move the robot by applying
very low external forces.
A passivity-based controller, similar to eq. (17) of the rigid
case, is obtained by choosing
u = τ
J,d
+ K
θ
(θ
d
θ) D
θ
˙
θ, (24)
with positive definite, diagonal matrices K
θ
and D
θ
. The
feedforward joint torque τ
J,d
and the motor reference θ
d
can be directly computed from the desired link trajectory
q
d
(t) (see, e.g., [22]). We note that the passivity properties
are preserved provided that the following condition on the
introduced (diagonal) damping matrices is satisfied [18]:
D
1
4
(D
s
D)
T
D
θ
(D
s
D).
Combining eqs. (23) and (24) yields the overall controller
τ = K
P
(θ
d
θ)K
D
˙
θ+K
P τ
(τ
J,d
τ
J
)K
˙
τ
J
+τ
J,d
,
(25)
where the expressions of the final gain matrices can be easily
obtained from the previous definitions. Equation (25) repre-
sents a full-state feedback with respect to suitable, possibly
time-varying, reference values. Having an explicit joint torque
feedback in addition to a position feedback loop, the above
control law can be parametrized by a suitable choice of its
gains so as to yield a position, a torque, or an impedance
controller [23]. This fact will be used for implementing
different strategies of robot reaction to collisions in a single
framework.
Realizing a zero-gravity condition for an elastic joint robot,
with a choice equivalent to eq. (16) of the rigid case, is
not as immediate. Adding just g(q) in the expression of τ
is not enough to completely eliminate gravity effects from
the picture. Rather, it can be shown that the more complex
dynamic term g(q)+BK
1
¨
g(q) would be needed. However,
it is often sufficient to compensate gravity in any static con-
figuration. This can be more easily achieved by adopting the
iterative scheme of [24] where, for each motor measurement
θ, a gravity term
¯
g(θ) is computed such that
¯
g(θ) = g(q), (θ, q) := {(θ, q)| K(θ q) = g(q)}.
This approach fits well within the passivity framework of
the controller. Another feasible scheme for on-line (partial)
compensation of gravity has been proposed in [9].
1627

Citations
More filters
Journal ArticleDOI
TL;DR: The present atlas is a result of the EURON perspective research project “Physical Human–Robot Interaction in anthropic DOMains (PHRIDOM)”, aimed at charting the new territory of pHRI, and constitutes the scientific basis for the ongoing STReP project ‘Physical Human-Robots Interaction: depENDability and Safety (PHRIENDS’.

699 citations


Cites methods from "Collision Detection and Safe Reacti..."

  • ...In [55], a different strategy has been implemented on a lightweight robot arm, by determining a direction of safe postimpact motion for the robot from the same signal used for collision detection....

    [...]

Journal ArticleDOI
TL;DR: An extensive review on human–robot collaboration in industrial environment is provided, with specific focus on issues related to physical and cognitive interaction, and the commercially available solutions are presented.

632 citations

Journal ArticleDOI
TL;DR: The first systematic experimental evaluation of possible injuries during robot‐human crashes using standardized testing facilities is presented, and a consistent approach for using these sensors for manipulation in human environments is described.
Abstract: – The paper seeks to present a new generation of torque‐controlled light‐weight robots (LWR) developed at the Institute of Robotics and Mechatronics of the German Aerospace Center., – An integrated mechatronic design approach for LWR is presented. Owing to the partially unknown properties of the environment, robustness of planning and control with respect to environmental variations is crucial. Robustness is achieved in this context through sensor redundancy and passivity‐based control. In the DLR root concept, joint torque sensing plays a central role., – In order to act in unstructured environments and interact with humans, the robots have design features and control/software functionalities which distinguish them from classical robots, such as: load‐to‐weight ratio of 1:1, torque sensing in the joints, active vibration damping, sensitive collision detection, compliant control on joint and Cartesian level., – The DLR robots are excellent research platforms for experimentation of advanced robotics algorithms. Space and medical robotics are further areas for which these robots were designed and hopefully will be applied within the next years. Potential industrial application fields are the fast automatic assembly as well as manufacturing activities done in cooperation with humans (industrial robot assistant). The described functionalities are of course highly relevant also for the potentially huge market of service robotics. The LWR technology was transferred to KUKA Roboter GmbH, which will bring the first arms on the market in the near future., – This paper introduces a new type of LWR with torque sensing in each joint and describes a consistent approach for using these sensors for manipulation in human environments. To the best of one's knowledge, the first systematic experimental evaluation of possible injuries during robot‐human crashes using standardized testing facilities is presented.

585 citations


Cites background from "Collision Detection and Safe Reacti..."

  • ...Inputs to the observer are the joint torque s and motor positions [5]....

    [...]

Proceedings ArticleDOI
14 Oct 2008
TL;DR: The proposed collision detection and reactions methods prove to work very reliably and are effective in reducing contact forces far below any level which is dangerous to humans.
Abstract: In the framework of physical human-robot interaction (pHRI), methodologies and experimental tests are presented for the problem of detecting and reacting to collisions between a robot manipulator and a human being. Using a lightweight robot that was especially designed for interactive and cooperative tasks, we show how reactive control strategies can significantly contribute to ensuring safety to the human during physical interaction. Several collision tests were carried out, illustrating the feasibility and effectiveness of the proposed approach. While a subjective ldquosafetyrdquo feeling is experienced by users when being able to naturally stop the robot in autonomous motion, a quantitative analysis of different reaction strategies was lacking. In order to compare these strategies on an objective basis, a mechanical verification platform has been built. The proposed collision detection and reactions methods prove to work very reliably and are effective in reducing contact forces far below any level which is dangerous to humans. Evaluations of impacts between robot and human arm or chest up to a maximum robot velocity of 2.7 m/s are presented.

544 citations


Cites background or methods from "Collision Detection and Safe Reacti..."

  • ...In particular, they will serve for a new method of scaling time increments in the trajectory generation, which allows the user to push the robot intuitively forth and back along its desired path even though the robot is still under position control....

    [...]

  • ...Furthermore, it will be showcased how the collision detection and reaction can help to prevent damage to the robotic structure and thus additionally contribute to an increase in safety due to fault protection....

    [...]

  • ...The collision detection mechanism proposed in [10], which provides a filtered version of the external collision torque τext, along with improvements concerning detection sensitivity and alternative detection schemes are presented and compared to each other....

    [...]

Journal ArticleDOI
TL;DR: This survey paper review, extend, compare, and evaluate experimentally model-based algorithms for real-time collision detection, isolation, and identification that use only proprioceptive sensors that cover the context-independent phases of the collision event pipeline for robots interacting with the environment.
Abstract: Robot assistants and professional coworkers are becoming a commodity in domestic and industrial settings. In order to enable robots to share their workspace with humans and physically interact with them, fast and reliable handling of possible collisions on the entire robot structure is needed, along with control strategies for safe robot reaction. The primary motivation is the prevention or limitation of possible human injury due to physical contacts. In this survey paper, based on our early work on the subject, we review, extend, compare, and evaluate experimentally model-based algorithms for real-time collision detection, isolation, and identification that use only proprioceptive sensors. This covers the context-independent phases of the collision event pipeline for robots interacting with the environment, as in physical human–robot interaction or manipulation tasks. The problem is addressed for rigid robots first and then extended to the presence of joint/transmission flexibility. The basic physically motivated solution has already been applied to numerous robotic systems worldwide, ranging from manipulators and humanoids to flying robots, and even to commercial products.

467 citations


Cites background or methods from "Collision Detection and Safe Reacti..."

  • ...The monitoring method based on the generalized momentum observer introduced in [33], [34], and [44] was motivated by the desire of avoiding the inversion of the robot inertia matrix, decoupling the estimation result, and also eliminating the need of an estimate of joint accelerations....

    [...]

  • ...The first method that achieved simultaneously collision detection, isolation, and identification was proposed in [33] and [34]....

    [...]

  • ...The following approach is an extension of our previous works [34], [40] by systematically deriving a threshold with higher robustness against model uncertainties and disturbances A collision detection function cd(....

    [...]

  • ...We start from our early works [34], [35], [40], [43], [44] and take advantage of the extensive experience gained over the years in developing, using, and refining our original methods....

    [...]

  • ...The monitoring signal r is also called residual vector (see [34] and [35])....

    [...]

References
More filters
Journal ArticleDOI

1,542 citations


"Collision Detection and Safe Reacti..." refers background in this paper

  • ...With the standard assumptions in [19], the dynamic model becomes...

    [...]

Journal ArticleDOI
TL;DR: In this paper, the authors considered the problem of designing joint-actuation mechanisms that can allow fast and accurate operation of a robot arm, while guaranteeing a suitably limited level of injury risk.
Abstract: This article considered the problem of designing joint-actuation mechanisms that can allow fast and accurate operation of a robot arm, while guaranteeing a suitably limited level of injury risk. Different approaches to the problem were presented, and a method of performance evaluation was proposed based on minimum-time optimal control with safety constraints. The variable stiffness transmission (VST) scheme was found to be one of a few different possible schemes that allows the most flexibility and potential performance. Some aspects related to the implementation of the mechanics and control of VST actuation were also reported.

620 citations

Proceedings ArticleDOI
18 Apr 2005
TL;DR: An implementation of such concepts, consisting of a novel electromechanical Variable Stiffness Actuation (VSA) motor, is described, along with experimental results showing performance and safety of a one-link arm actuated by the VSA motor.
Abstract: This paper is concerned with the design and control of actuators for machines and robots physically interacting with humans, implementing criteria established in our previous work [1] on optimal mechanical-control co-design for intrinsically safe, yet performant machines. In our Variable Impedance Actuation (VIA) approach, actuators control in real-time both the reference position and the mechanical impedance of the moving parts in the machine in such a way to optimize performance while intrinsically guaranteeing safety. In this paper we describe an implementation of such concepts, consisting of a novel electromechanical Variable Stiffness Actuation (VSA) motor. The design and the functioning principle of the VSA are reported, along with the analysis of its dynamic behavior. A novel scheme for feedback control of this device is presented, along with experimental results showing performance and safety of a one-link arm actuated by the VSA motor.

543 citations


"Collision Detection and Safe Reacti..." refers background in this paper

  • ...Also, variable stiffness actuation can be used, stiffening the joints during low-velocity transients and relaxing them at high-velocity regimes [10], in such a way that tracking performance during free motion is not compromised by the introduced compliance....

    [...]

Journal ArticleDOI
TL;DR: By drastically reducing the effective impedance of the manipulator while maintaining high-frequency torque capability, it is shown that the competing design requirements of performance and safety can be successfully integrated into a single manipulation system.
Abstract: In recent years, many successful robotic manipulator designs have been introduced. However, there remains the challenge of designing a manipulator that possesses the inherent safety characteristics...

466 citations

Proceedings ArticleDOI
18 Apr 2005
TL;DR: The idea is to handle a collision at a generic point along the robot as a fault of its actuating system as well as a previously developed dynamic FDI technique, which does not require acceleration or force measurements.
Abstract: We consider the problem of real-time detection of collisions between a robot manipulator and obstacles of unknown geometry and location in the environment without the use of extra sensors. The idea is to handle a collision at a generic point along the robot as a fault of its actuating system. A previously developed dynamic FDI (fault detection and isolation) technique is used, which does not require acceleration or force measurements. The actual robot link that has collided can also be identified. Once contact has been detected, it is possible to switch to a suitably defined hybrid force/motion controller that enables to keep the contact, while sliding on the obstacle, and to regulate the interaction force. Simulation results are shown for a two-link planar robot.

332 citations


"Collision Detection and Safe Reacti..." refers background or methods in this paper

  • ...In this paper, we build upon the original idea of [15] 1-4244-0259-...

    [...]

  • ...In fact, no possible robot motion would be able to reduce the force loading in this case....

    [...]

  • ...A detection scheme that works under similar conditions and avoids the above drawbacks has been recently proposed in [15]....

    [...]

  • ...Note that σ can be computed using the measured joint position q and velocity q̇ (possibly obtained through numerical differentiation) and the commanded motor torque τ ....

    [...]

Frequently Asked Questions (1)
Q1. What have the authors contributed in "Collision detection and safe reaction with the dlr-iii lightweight manipulator arm" ?

Based on physical quantities such as total energy and generalized momentum of the robot manipulator, the authors present an efficient collision detection method that uses only proprioceptive robot sensors and provides also directional information for a safe robot reaction after collision. Experimental results on collisions with the DLR-III lightweight manipulator are reported.