Beruflich Dokumente
Kultur Dokumente
Abstract
In this article, we develop and assess a novel approach for the control of underactuated planar bipeds that is based on
velocity decomposition. The new controller employs heuristic rules that mimic the functionality of transverse linearization
feedback control and that can be layered on top of a conventional hybrid zero dynamics (HZD)-based controller. The
heuristics sought to retain HZD-based control’s simplicity and enhance disturbance rejection for practical implementation
on realistic biped robots. The proposed control strategy implements a feedback on the time rate of change of the decom-
posed uncontrolled velocity and is compared with conventional HZD-based control and transverse linearization feedback
control for both vanishing and non-vanishing disturbances. Simulation studies with a point-foot, three-link biped show
that the proposed method has nearly identical performance to transverse linearization feedback control and outperforms
conventional HZD-based control. For the non-vanishing case, the velocity decomposition-enhanced controller outper-
forms HZD-based control, but takes fewer steps on average before failure than transverse linearization feedback control
when walking on uneven terrain without visual perception of the ground. The findings were validated experimentally on a
planar, five-link biped robot for eight different uneven terrains. The velocity decomposition-enhanced controller outper-
formed HZD-based control while maintaining a relatively low specific energetic cost of transport (;0.45). The biped
robot ‘‘blindly’’ traversed uneven terrains with changes in terrain height accumulating to 5% of its leg length using the
stand-alone low-level controller.
Keywords
Underactuated robots, legged locomotion, velocity decomposition, disturbance rejection, uneven terrain walking
encounter both types of disturbance, sometimes at the same evaluate the stochastic performance of dynamic walkers is
time, and low-level feedback controllers must be able to varying slopes (Byl and Tedrake, 2009; Saglam and Byl,
reject both. As such, the proposed control method should 2018). Here, the change in height Dh on each step is ran-
yield disturbance rejection performance that is consistent domly generated from a normal Gaussian distribution
across a wide range of disturbance for it to be a viable N (m, s), where the average ground height m is set to zero
terrain-blind strategy. and the standard deviation of the distribution s loosely cor-
Since most notions of stability rely on conservative con- responds to the roughness of the terrain (Bledt et al., 2018).
straints that do not necessarily predict what happens in the Therefore, when comparing stochastic performance, the
real world, this section describes practical metrics of perfor- controller that maintains stability with the higher s will be
mance for both vanishing and non-vanishing disturbances. referred to as the more robust one. The H2 and H‘ norms
For example, violating friction constraints at some particu- (Fan and Manchester, 2018; Hobbelen and Wisse, 2007)
lar instants of the gait cycle does not necessarily mean that both assume finite energy and fail to practically predict
the robot will fall (Saglam and Byl, 2018). Here, we loosely when one should expect the robot to fall. The step height
define stable walking as the successive repetition of multi- metric is directly observable and can be translated from
ple steps (Yang et al., 2009). simulation to experimental studies, and vice versa.
If a single, finite velocity disturbance does not drive the The mean of first passage time quantifies the expected
system dynamics out of the basin of attraction, classical time to failure (Byl and Tedrake, 2009). It is not, however,
definitions of stability for periodic walking can be applied computationally feasible for complex models of walking
to quantify the speed of return to the target motion. Control because, based on a Markov chain model, it requires dis-
performance can be assessed by the maximum magnitude cretization of the entire state space. The average number of
of the disturbance that can be rejected, as well as the speed steps taken before failure for a given level of terrain noise
and energetic cost of recovery. The speed of recovery can provides similar information, so it is the metric of the prob-
be quantified by the time needed to return to within some ability of falling used herein.
bounds of the zero dynamics trajectory or the number of In the case of non-vanishing disturbances, when the
steps required to return within some limits of the nominal dynamics drift around the state space, probability density
step velocity. The latter, discrete metric is used throughout functions (PDFs) quantify a controller’s ability to follow the
this article for consistency with the comparison of the ener- desired gait. In simulation and experiment, the step velocity
getic efficiency of recovery, measured via the specific ener- PDF is used to compare the controllers’ ability to maintain
getic cost of transport (SCOT). The SCOT is a dimensionless the desired walking speed. Two-dimensional PDFs of the
measure quantifying the (total) energy required to move one unactuated dynamics are also used to compute the meta-
kilogram one meter (Collins et al., 2005). Energetic efficiency stable neighborhoods of state space, which characterize the
undeniably plays a critical role in legged robotics, so the velo- behavior and attractiveness of the fixed point of the dyna-
city decomposition-enhanced controller should enable gaits mical system under different controllers in the presence of
with relatively low SCOT, and the trade-offs between robust- stochastic perturbations (Byl and Tedrake, 2009).
ness and efficiency are analyzed for each type of disturbance
investigated in this work.
To compare the magnitude of the disturbance that each
3. Mathematical framework
controller can reject at different points of the gait cycle, the 3.1. Hybrid dynamical system
state-space region of stability of the unactuated dynamics
The dynamics of bipedal locomotion can be modeled as a
is computed numerically. For the point-foot dynamic walk-
series of finite-time single support phases and infinitesi-
ers underactuated by one in this article, this is a two-
mally short ground impacts. In single support, the swing
dimensional projection of the basin of attraction onto the
leg moves from behind to in front of the stance leg. A foot
unactuated dynamics. Computing the basin of attraction of
impact transfers support from the stance to the swing leg,
complex biped robots can become intractable as more
swapping their roles and redirecting the velocity of the
DOFs are added, which motivates the use of simpler alter-
biped’s CoM.
natives such as the unactuated dynamics.
The two biped models in this article consist of three and
For the second type of vanishing perturbations, the max-
five rigid links. More generally, for a planar, left–right sym-
imum height of a single stair that the robot can surmount
metric biped with n rigid links, single support is modeled using
and continue walking is the metric of comparison for the
continuous second-order ordinary differential equations,
controllers. Since this measure is inherently related to leg
length, all the changes in terrain height are reported as per- _ q)q_ + G(q) = B(q)u
M(q)€q + C(q, ð1Þ
centages of leg length.
When the biped robot walks on uneven terrain, the per- where q = ½q1 , q2 , . . . , qn 2 Q R is a set of general-
n
sistent disturbances can cause the robot to fall down or ized coordinates that fully defines the position of the biped
cause the system dynamics to be pushed away from the zero in space. M(q) 2 Rn × n is the symmetric inertia matrix,
dynamics. For this work, uneven terrains are modeled using _ 2 Rn × n is the Coriolis matrix, G(q) 2 Rn is the
C(q, q)
stairs of varying heights. An equally complex approach to gravity vector, B(q) 2 Rn × m is the actuation matrix
Fevre et al. 5
mapping actuator torques to joint torques, and u 2 U Rm approach is to choose the constraints to be smooth func-
is a vector of independent control inputs. By definition, an tions of position only, which for this work are chosen as
n-dimensional system is underactuated if it has fewer than Bézier polynomials. The vector of functions
n actuators (m\n). The actuation matrix varies based on F(u) = (f1 (u), . . . , fm (u))T represents the desired config-
the level of actuation of the system. If the biped is underac- uration of the biped during a given step, and the output
tuated, the rank of the actuation matrix is less than the functions are defined as the tracking errors for each actu-
number of DOFs, and the system is not full-state feedback ated DOF,
linearizable. The models analyzed herein assume no ankle
actuation, meaning that the angle between the ground and y = h(q) = qa F(u) ð6Þ
the stance leg tibia is not directly controlled. All other
where qa = (q1 , . . . , qm )T is the vector of actuated
joints of the robot are actuated. Hence, for this work,
coordinates.
rank(B(q)) = n 1.
For walking on flat ground, the values of the phase vari-
The hypersurface S is a switching set that defines the
able at the beginning and end of every step can be written
limits of the continuous dynamics and relates the general-
ized coordinates q immediately before impact to the state as u+ and u, respectively. In other words, u+ and u corre-
q+ immediately after impact via the discrete map, spond to 0 and 100% of the gait cycle, in which the Bézier
polynomials are well-defined. When the robot is walking
q+ = Dq q , q 2 S on uneven terrain without ground perception, however, the
ð2Þ phase variable takes on values that are outside of its pre-
q_ + = Ddq q_ , q 2 S defined interval. For this work, both in simulation and
where Dq is the n × n switching matrix and Ddq is an n × n experiment, u was saturated to stay between 0 and 100% of
matrix relating pre- and post-impact velocities. Here Dq can the gait cycle such that
be obtained by inspection even with complex kinematic
chains, and Ddq can be derived via conservation of angular fi (u) = fi (u ), 8u ø u
ð7Þ
momentum about the point of ground contact (Westervelt fi (u) = fi (u+ ), 8u ł u+
et al., 2007). The two phases of a step result in a hybrid
model for walking, and Equations (1)–(2) form the system’s Although this method is suitable for dealing with modest
full hybrid dynamics. changes in terrain height (Dh;5% of leg length), for more
The guard set of the hypersurface S changes at every significant stairs, a different method to cope with Bezier
step, and the impact map is applied when polynomials outside of the (u+ , u ) range should be
developed.
S = fpy : Rn ! R j py = Dh, p_ y \0g ð3Þ
where py and p_ y are the vertical position and velocity of the 3.2. Velocity decomposition
swing foot, respectively. Note that Equation (3) still holds This section derives the relevant terms for the proposed
for walking on flat ground (Dh = 0). velocity decomposition-enhanced controller. A thorough
For convenience, the hybrid system may be written in derivation of the velocity decomposition metric can be
state-space form using coupled first-order differential found elsewhere in the literature (Goodwine and
equations Nightingale, 2010; Nightingale et al., 2008).
( Velocity decomposition is applicable to all underactu-
x_ = f(x) + g(x)u, x 62 S ated mechanical systems that can be described by a
S: ð4Þ
x = D(x ),
+
x2S Lagrangian of the form L(q, q) _ = 12 q_ T M(q)q_ V(q),
1 _T
where 2 q M(q)q_ is the kinetic energy and V(q) is the
where x = ½qT , q_ T T 2 X Rn defines the state space and potential energy. Inspired by geometric analysis and con-
trol, the velocity decomposition metric uses differential
q_ geometry to partition the equations of motion of an under-
f(x) = 1
M(q) (C(q, _ _
q)q + G(q)) actuated mechanical system into directions aligned with the
ð5Þ
0 inputs, termed controlled directions, and directions ortho-
g(x) =
M(q)1 B(q) gonal to the inputs with respect to the inertia matrix,
termed uncontrolled directions.
A monotonically increasing phase variable u : Q ! R
The inverse of the inertia matrix is used to construct an
measures the progression of a step. While several options
Morthonormal basis with the linearly independent control
are physically valid, this work uses the unactuated angle
forces F i (i = 1, . . . , m),
between the upright vertical and the stance leg tibia for u.
The monotonic behavior of the phase variable makes it an
M1 F i
ideal candidate to parameterize a set of holonomic con- Yi = ð8Þ
straint functions on the m actuated coordinates. A common jjM1 F i jj2M
6 The International Journal of Robotics Research 00(0)
The set of vector fields Y = fY1 , . . . , Ym g 2 Rn × m repre- 1 kl ∂Mil ∂Mjl ∂Mij
Gkij = M + ð12Þ
sents the input distribution of the system, where jjjj2M 2 ∂qj ∂qi ∂ql
denotes the Euclidean norm with respect to M.
The M-orthonormal complement to the input distribu-
n × (nm) 4. Control strategies
tion is represented by Y ? = fY? 1 , . . . , Ynm g 2 R
?
For mechanical systems underactuated by one, the K(t) = R1 B(t)T P(t), where P(t) is the solution to the
dynamics of the unactuated coordinate u are given by the JMRE. A known transformation given in Manchester et al.
zero dynamics equation (2011) takes the time-based controller gains to state-based
gain scheduling and
a(u)€u + b(u)u_ 2 + g(u) = 0 ð17Þ
K(u) = R1 B(u, u)
_ T P(u) ð23Þ
where a(), b(), and g() are given in Appendix B. The
zero dynamics equation has a general integral of motion proves to locally exponentially orbitally stabilize the origi-
h i nal nonlinear system. The complexity of the optimal control
R
_ = u_ 2 c(u0 , u) u_ 2 u c(s, u0 ) 2g(s) ds
I(u, u) problem, however, has limited its application to point-foot
n 0 R u0 o a(s) ð18Þ robots with no more than two DOFs.
u b(t)
c(u0 , u) = exp 2 u0 a(t) dt
which preserves its value along the solution, given (u0 , u_ 0 ) 4.3. Velocity decomposition-enhanced control
are chosen on the target trajectory. This makes I an ideal
The deviation from the periodic orbit is computed in terms
candidate for the additional state quantifying deviation from
of the error in the derivative of the decomposed uncon-
the desired orbit. Therefore, the complete set of transverse
trolled velocity ds
dt ,
coordinates becomes x? = ½I, yT , y_ T T . In a small enough
neighborhood of the periodic orbit, the linearization of the d(es ) d
continuous transverse dynamics at time t is given by = (sref s) ð24Þ
dt dt
z_ = A(t)z(t) + B(t)v(t), t 6¼ tj ð19Þ Simulation studies and experimental work showed that
measuring deviations from the periodic orbit throughout the
where A( ) and B( ) are given in Manchester et al.
gait cycle using the novel differential coordinates provides
(2011), and tj, j = 1, 2, . . . , are the times of ground
superior disturbance rejection performance (Fevre et al.,
impacts. The transverse linearization of impacts is given by
2018). The idea that the heuristic controller is locally equiv-
z(t+ ) = Fz(t ) ð20Þ alent to transverse linearization feedback control, however,
relies more on simulations and experimental validations,
where F is given in Shiriaev et al. (2008). Using a periodic rather than a well-established proof. For control purposes,
matrix gain K(t) = K(t + T ), the state feedback controller proportional gains define the torso and leg offsets
s)
v = K(t)z ð21Þ Toffset = KT d(e
dt
s)
ð25Þ
exponentially stabilizes Equation (19) (Shiriaev et al., Loffset = KL d(e
dt
2005), which is equivalent to exponential orbital stabiliza-
The change in the desired torso pitch angle, Toffset , pri-
tion of Equation (17)
marily affects the gravity contribution to dsdt . In a decelera-
A receding-horizon control strategy can be used to
tion case, the error is positive, so the positive Toffset shifts
implement transverse linearization feedback control in real-
the biped’s CoM forward to regain momentum. In an accel-
time and guarantee orbital stability of the nonlinear system
eration case, the reverse applies, as the negative torso offset
to the periodic target (x? ! 0). A receding-horizon optimal
shifts the CoM backward to reduce momentum. Similarly,
control is obtained by computing the transverse lineariza-
the change in the desired swing leg angle, Loffset , acts to
tion about the periodic orbit a few steps ahead (Manchester
swing the leg forward in a deceleration case and acts to drag
et al., 2011). The optimal control problem is solved by
the leg in an acceleration case.
minimizing the cost function
The offsets are layered on top of the hi (u) output func-
Z tf tions of the HZD-based controller from Section 4.1. The
J (x, u) = ½z(t)T Q(t)z(t) + v(t)T R(t)v(t)dt heuristics were derived from the three-link biped model
ti shown in Figure 2, which is identical to that in Westervelt
Nj
X ð22Þ
T
et al. (2007), where q1 and q2 are actuated and q3 is unac-
+ z(tj ) Qj z(tj ) tuated. The legs each have a length of 1 m and a 5 kg point
j=1
mass halfway between the foot and hip. There is a 15 kg
which is similar to a constrained linear quadratic regulator point mass at the hip, in addition to the 10 kg point mass at
(LQR) problem, where Q, R, and Qj are weighting func- the tip of the torso, which is 0.5 m long. The resulting con-
tions on the states, inputs, and step-end positions, respec- trol law for the velocity decomposition-enhanced controller
tively. The optimal control is computed by solving the becomes
jump-matrix Riccati equation (JMRE) backward in time
from tf to ti for a constant number of footsteps ahead. The q1 (f1 (u) + Toffset )
h(q) = ð26Þ
solution is then used to construct the matrix of gains q2 (f2 (u) + Loffset )
8 The International Journal of Robotics Research 00(0)
Table 2. Simulated SCOT versus gait speed for curved and Table 3. Maximum height of a single stair that the robot could
point-foot gaits. clear for each control strategy.
Gait speed (m/s) Point-foot SCOT Curved-foot SCOT HZD Vel Dec TLFC
average number of steps taken before failure effectively Fig. 4. Simulated number of steps taken before failure for
measures the stochastic performance of each controller. different levels of terrain noise. The circles represent the average
Once again, the 0.75 m/s point-foot gait was evaluated over for each controller, and the funnels are drawn using the standard
each of the 100 terrain profiles for each noise level. The deviation (peak-to-peak = 1 standard deviation).
three-link model started each trial at steady state and ended
after completing the 25,000 steps or falling down. The
results are summarized in Figure 4, and a side-by-side video
comparison of HZD-based control and velocity decomposi-
tion-enhanced control is given in Extension 1. HZD-based
control reached the 25,000 step milestone 100 times in a
row for s = 0:858, whereas velocity decomposition-
enhanced control and transverse linearization feedback con-
trol did so for s = 1:558 and s = 1:808, respectively. This
corresponds to the controllers successfully completing
2,500,000 steps in a row on a random terrain at their respec-
tive noise levels.
Next, the controllers’ ability to maintain a desired walk-
ing speed was assessed using the step velocity PDF for 100
terrains, each consisting of 30 steps for s = 0:808. This Fig. 5. Step velocity PDF of 3,000 steps when terrain height
level of noise permitted the controllers to complete each noise s = 0:88 using the 0.75 m/s point-foot gait.
trial and keep the states within the regions of stability. The
robot once again started each simulation on the zero Table 4. Step velocity distribution and SCOT for 3,000 steps
dynamics trajectory, and kinematic data for all 3,000 steps when s = 0:808. The step velocity PDF is shown in Figure 5. The
were recorded. The step velocity PDF was computed for design speed is 0.75 m/s.
each controller and is reported in Figure 5, which shows
HZD Vel Dec TLFC
that velocity decomposition-enhanced control was able to
better maintain walking at the design speed than HZD- Step velocity (m/s) 0.73 6 0.07 0.74 6 0.05 0.74 6 0.04
based control. Transverse linearization feedback control, SCOT 0.150 0.151 0.148
however, outperformed both control strategies. Energetic t 1, max (Nm) 152.7 153.9 151.5
efficiency measures are also reported in Table 4. The t 2, max (Nm) 105.0 106.7 105.6
SCOT and the peak torques were similar for all three con-
trollers, so the enhanced robustness did not come at the
expense of efficiency for the latter two controllers. and 99% of the data for the 3,000 steps (Figure 6). If the
The attractiveness of a two-link walker’s fixed point at three-link model has not fallen as t ! ‘, the states will most
impact was shown to have a positive effect on the perfor- likely be in those regions as they represent the regions of the
mance of the biped on stochastic terrain (Byl and Tedrake, state space into which the biped is being pulled. The approach
2009). While the stability of a fixed point can be assessed is analogous to computing the PDF in two dimensions.
for periodic walking, the fixed point’s attractiveness at The top graph in Figure 6 corresponds to HZD-based
impact for uneven terrain can be measured by computing control, whereas the middle and bottom graphs correspond
metastable neighborhoods of the state space that the to the velocity decomposition-enhanced controller and
dynamics visit more or less often. The metastable neigh- transverse linearization feedback control, respectively.
borhoods of (u, u) _ were computed to capture 50%, 90%, Because the stochastic performances of velocity
Fevre et al. 11
Steps to HZD 9 9 9 9
steady state Vel Dec 5 5 6 5.3
approximately 600 steps (Fevre et al., 2018). The velocity 6.3.2. Uneven terrain. The controllers were compared in
decomposition-enhanced controller had reduced variance terms of the maximum height of a single wooden stair that
in step velocity for the undisturbed experiments, showing the robot could surmount. To be considered a successful
that the proposed control method produced more consistent trial, the robot must have been able to return to steady state
steady-state walking and dealt with model uncertainties following the disturbance. This disturbance was considered
better than HZD-based control. a vanishing perturbation because it was modeled as a single
Because the heuristic rules are layered on top of the elevated stair such that the next step involved a return to
HZD-based formulation, velocity decomposition-enhanced the standard terrain height. Because the robot has no terrain
control is identical to HZD-based control when the error in preview to adjust the swing leg trajectory and avoid obsta-
ds
dt is zero at steady state. This is true in simulation (Table 2) cles, the steps were positioned in the room to avoid toe-
and almost true in experiment given a low noise level. scuffing such that the robot’s swing foot would clear the
More steady-state experiments were conducted (in addition step and impact the top surface of the step. Mathematically,
to those from Fevre et al. (2018)) using slower and faster this is equivalent to
gaits to verify that velocity decomposition-enhanced con-
trol did not expend more energy when regulating the error Dh = Lt (cos(q1 + q3 + q5 ) + cos(q2 + q4 + q5 ))
in ds
dt . The SCOT of both controllers is plotted as a function
ð29Þ
+ Lf (cos(q1 + q5 ) + cos(q2 + q5 ))
of step velocity in Figure 8, which shows that the energetic
efficiency of HZD-based control is retained under velocity always being true at impact, where Lt and Lf are the lengths
decomposition-enhanced control despite measurement of the tibia and femur, respectively. Equation (29) corre-
errors. sponds to the hypersurface condition of Equation (3) for
ERNIE. When this condition is met, the instantaneous
transfer of support occurs, and the next step is initiated.
6.3. Vanishing perturbations Figure 7 illustrates this idea.
6.3.1. Velocity disturbances. The velocity disturbance For all experiments, a 0.70 m/s point-foot gait optimized
experiments reported in Fevre et al. (2018) implemented for SCOT from Brown and Schmiedeler (2016) was used.
perturbations repeatedly using a barrier that would hit the This is the gait used with ERNIE in Fevre et al. (2018), for
distal end of the boom and decelerate the robot. The results which the robot walked reliably. Velocity decomposition-
are summarized in Table 6. The experimental results vali- enhanced control was able to clear stairs that were 5.2% of
date the simulation findings, showing that the proposed the robot’s leg length three times in a row, whereas HZD-
formulation improved the robot’s ability to reject a velocity based control only cleared stairs up to 3.8% three times in
disturbance in experiment, just as in simulation. Note that, a row. Thus, the proposed control method improved the
in experiment, the velocity disturbance was not a true disturbance rejection performance to both types of vanish-
impulse as in simulation. The SCOT reported in Table 6 ing perturbations considered in experiment: velocity distur-
corresponds to the average SCOT of all the steps between bances and walking on uneven terrain.
the time of the disturbance and the time of return to steady
state. As expected, the SCOT to return to steady state in 9
steps for the HZD-based controller was lower than for the 6.4. Non-vanishing perturbations
velocity decomposition-enhanced controller that returned To compare the stochastic performance of the controllers
to steady state nearly twice as fast, in 9 steps. Yet, a SCOT experimentally, terrains of variable heights were
Fevre et al. 13
Fig. 10. The steps were placed such that the robot did not have the ability to return to steady state before the next disturbance was
encountered. Terrains 6 and 7 are qualitatively similar, but the stochasticity of the terrain was increased for terrain 7.
14 The International Journal of Robotics Research 00(0)
Terrain HZD Vel Dec HZD Vel Dec HZD Vel Dec
1 ü ü ü ü ü ü
2 ü ü ü ü ü ü
3 3 ü 3 ü 3 ü
4 3 4 3 4 3 4
5 4 ü 4 ü 4 ü
Finished
trials 2 4 2 4 2 4 Fig. 11. Step velocity PDF of 160 inside steps for modest
changes in terrain heights using a 0.70 m/s point-foot gait.
Non-vanishing perturbation cases involved walking on bipeds, and the approach was used to produce highly
uneven terrain. The velocity decomposition-enhanced con- dynamic walking motions in three dimensions. Therefore,
troller clearly outperformed HZD-based control and closely although the velocity decomposition-enhanced controller
approximated transverse linearization feedback control in was not tested on spatial bipeds, the approach is not limited
terms of the number of steps taken before failure in simula- to planar robots since the formulation is, at its core, an
tion. The experimental results were consistent with the HZD-based controller. Moreover, since transverse lineariza-
simulations except that, owing to the complexity of the tion feedback control is applicable to several degrees of
optimal control problem, transverse linearization feedback underactuation (Shiriaev et al., 2010), additional heuristic
control could not be implemented on the five-link biped. rules can be derived to improve disturbance rejection per-
Whereas earlier work by the authors tracked the error in formance in the frontal plane for improved lateral
hip velocity (equivalent to the unactuated velocity u_ in con- stabilization.
ventional coordinates) to compute the heuristics, this work The robustness of the velocity decomposition-enhanced
uses novel coordinates inspired by differential geometry to controller to a wide range of disturbance also motivates the
implement a feedback on ds dt and increase the robustness of need for deriving simple heuristics from formal control
underactuated biped robots to disturbances in their unactu- methods for practical applications on realistic bipeds. In
ated DOFs. Previous work showed that measuring devia- comparison with control strategies that rely on quasi-static
tions from the zero dynamics trajectory using dsdt as opposed stability principles, the proposed method enables dynamic
to u_ provides superior robustness and faster return to steady human-like gaits with relatively low SCOT within the HZD
state, which can be explained in part by two factors. framework and improves their robustness to bridge the gap
between robust bipedal locomotion and energetic effi-
s)
1. A simple proportional feedback on d(e
dt corresponds to
ciency. While terrain perception must become a necessity
a time-varying controller with state-based gain sche- for biped robots to achieve their envisioned potential in
duling since the direction of the input vector field today’s society, the proposed terrain-blind control strategy
changes throughout the gait (Section 5.1). replicates the worst-case scenario by simulating long-term
2. The time-varying controller resembles the gain tasks in which sensing errors and/or reported changes in
scheduling method employed by transverse lineariza- terrain heights accumulate to 5% of the robot’s leg length.
tion feedback control to guarantee orbital stability
(Section 4.2).
7.2. Limitations due to low coupling
It is clear from Equation (10) that the control inputs do not
appear in the derivative of the decomposed uncontrolled
7.1. Broader applications velocity s. Thus, the only way the control can influence the
In addition to providing a powerful analytical tool for uncontrolled dynamics is through the coupling terms. The
designing dynamic walking gaits and controllers, the expression for how s changes with time provides an analy-
method of HZD has also been extended to the design of tical measure of the dynamic coupling between the con-
running motions. HZD-based control was later extended to trolled and uncontrolled velocities. The practical
allow for periods of multi-contact, making the formulation implication is that the measure of coupling quantifies the
not only applicable to underactuated systems but to fully instantaneous control authority over the unactuated DOFs.
actuated humanoid robots with flat feet as well (Hereid Multiplying a w2 term, the B1 coefficients quantify how
et al., 2016). Another area of research in which HZD has much the controlled velocities can directly affect the rate of
contributed is that of robotic rehabilitation. Recent applica- change of the uncontrolled velocity. The B2 and B3 terms
tions employed the method for controlling an exoskeleton give a measure of the nonlinear coupling between the con-
for paraplegic individuals (Gurriet et al., 2018) and for trolled and uncontrolled velocities since they multiply both
developing lower-limb powered prostheses in human–robot w and s. The B4 coefficients multiply an s2 term, so these
shared control (Martin and Gregg, 2017). As such, terms are independent of the control inputs. Similarly, the
by augmenting the disturbance rejection performance B5 term represents the gravity contribution to the rate of
of conventional HZD-based control, the velocity change of the uncontrolled velocities.
decomposition-enhanced controller has the potential to When no coupling is available, the control inputs cannot
increase the practicality of a wide class of biped robots. help reject external disturbances in the uncontrolled DOFs
One advantage of the proposed approach is that the until the system moves away from the dynamic singularity
velocity decomposition metric is very general in that it and into a configuration in which there is coupling.
applies to any underactuated mechanical system that can be Conversely, when the instantaneous, relative amount of
described by Lagrange’s equations. Section 3.2 highlights coupling is strong, robustness to disturbances in the uncon-
that the decomposition holds for mechanical systems with trolled DOFs is higher. In the case of robotic bipedal loco-
several degrees of underactuation, a phenomenon intro- motion, however, high-bandwidth control is necessary to
duced by walking in 3D and flight phases in planar running achieve highly dynamic motions, and waiting for the robot
gaits. The HZD formulation was also expanded to spatial to move away from a dynamic singularity is not viable.
16 The International Journal of Robotics Research 00(0)
Previous work, however, suggests that the amount of cou- Chen T, Ni X, Schmiedeler JP and Goodwine B (2017) Using a
pling in most walking motions is low for underactuated nonlinear mechanical control coupling metric for biped robot
biped robots (Chen et al., 2017) control and design. In: IEEE International Conference on
Consequently, the error in ds Methods and Models in Automation Robotics, pp. 903–908.
dt is mainly affected by
changes in the terms that are independent of the control Chevallereau C, Grizzle JW and Shih CL (2009) Asymptotically
stable walking of a five-link underactuated 3D bipedal robot.
input(s), and correcting errors in the uncontrolled velocity
IEEE Transactions on Robotics 25(1): 37–50.
(or velocities) using the coupling terms alone becomes a Chew CM, Pratt J and Pratt G (1999) Blind walking of a planar
challenging task. As such, the feedback on the error in dsdt , bipedal robot on sloped terrain. In: IEEE International Confer-
and in general, the practicality of underactuated biped ence on Robotics and Automation, pp. 381–386.
robots, suffer from a lack of control authority, as seen Collins S, Ruina A, Tedrake R and Wisse M (2005) Efficient
through the coupling terms. Increasing the amount of cou- bipedal robots based on passive–dynamic walkers. Science
pling during the gait design process, however, helps solve 307(5712): 1082–1085.
this issue. Indeed, ongoing research by the authors suggests Da X, Harib O, Hartley R, Griffin B and Grizzle JW (2016) From
that optimizing the coupling in bipedal gaits enlarges their 2D design of underactuated bipedal gaits to 3D implementa-
tion: Walking with speed tracking. IEEE Access 4: 3469–3478.
regions of attraction and improves their stochastic perfor-
Da X, Hartley R and Grizzle JW (2017) Supervised learning for
mances. The results reported herein also highlight the need
stabilizing underactuated bipedal robot locomotion, with out-
for more formal ways of implementing control laws that door experiments on the wave field. In: IEEE International
make use of velocity decomposition for underactuated Conference on Robotics and Automation, pp. 3476–3483.
mechanical systems, which becomes possible when bipedal Fan F and Manchester I (2018) Robust control of dynamic walking
gaits are designed with increasing coupling. robots using transverse H‘ . In: IEEE International Conference
on Robotics and Automation, pp. 418–425.
Fevre M, Goodwine B and Schmiedeler JP (2018) Design and
Funding
experimental validation of a velocity decomposition-based con-
This work was supported by the National Science Foundation troller for underactuated planar bipeds. IEEE Robotics & Auto-
(grant number IIS-1527393). mation Letters 3(3): 1896–1903.
Fevre M, Goodwine B and Schmiedeler JP (2019) Velocity
decomposition-enhanced control for point and curved-foot pla-
ORCID iD
nar bipeds experiencing velocity disturbances. ASME Journal
Martin Fevre https://orcid.org/0000-0002-4238-5450 of Mechanisms and Robotics 11(2): 020901.
Goodwine B and Nightingale J (2010) The effect of dynamic sin-
gularities on robotic control and design. In: IEEE International
References
Conference on Robotics and Automation, pp. 5213–5218.
Alexander RM (1995) Simple models of human movement. Griffin B and Grizzle J (2017) Nonholonomic virtual constraints
Applied Mechanics Reviews 48(8): 461–470. and gait optimization for robust walking control. The Interna-
Banaszuk A and Hauser J (1995) Feedback linearization of trans- tional Journal of Robotics Research 36(8): 895–922.
verse dynamics for periodic orbits. Systems & Control Letters Gurriet T, Finet S, Boeris G, et al. (2018) Towards restoring loco-
26(2): 95–105. motion for paraplegics: Realizing dynamically stable walking
Bhounsule PA, Cortell J, Grewal A, et al. (2014) Low-bandwidth on exoskeletons. In: IEEE International Conference on
reflex-based control for lower power walking: 65 km on a sin- Robotics and Automation, pp. 2804–2811.
gle battery charge. The International Journal of Robotics Hereid A, Cousineau EA, Hubicki CM and Ames AD (2016) 3D
Research 33(10): 1305–1321. dynamic walking with underactuated humanoid robots: A
Bledt G, Wensing PM, Ingersoll S and Kim S (2018) Contact direct collocation framework for optimizing hybrid zero
model fusion for event-based locomotion in unstructured ter- dynamics. In: IEEE International Conference on Robotics and
rains. In: IEEE International Conference on Robotics and Automation, pp. 1447–1454.
Automation. Hobbelen DG and Wisse M (2007) A disturbance rejection mea-
Boston Dynamics Inc. (2009) Bigdog reflexes. https://youtu.be/ sure for limit cycle walkers: The gait sensitivity norm. IEEE
3gi6Ohnp9x8. Transactions on Robotics 23(6): 1213–1224.
Boston Dynamics Inc. (2013) Legged robot testing in desert. Hubicki C, Grimes J, Jones M, et al. (2016) Atrias: Design and
https://youtu.be/LJZQ3n-iQYE. validation of a tether-free 3D-capable spring-mass bipedal
Brown TL and Schmiedeler JP (2016) Reaction wheel actuation robot. The International Journal of Robotics Research 35(12):
for improving planar biped walking efficiency. IEEE Transac- 1497–1521.
tions on Robotics 32(5): 1290–1297. Isidori A (2013) Nonlinear control systems. New York: Springer
Buss BG, Ramezani A, Hamed KA, Griffin BA, Galloway KS Science & Business Media.
and Grizzle JW (2014) Preliminary walking experiments with Lee J, Kim JH and Oh Y (2016) A novel performance measure
underactuated 3D bipedal robot marlo. In: IEEE/RSJ Interna- for biped robots against bounded persistent disturbances. In:
tional Conference on Intelligent Robots and Systems, pp. IEEE/RSJ International Conference on Intelligent Robots and
2529–2536. Systems, pp. 5805–5812.
Byl K and Tedrake R (2009) Metastable walking machines. The Liu Y, Wensing PM, Schmiedeler JP and Orin DE (2016) Terrain-
International Journal of Robotics Research 28(8): blind humanoid walking based on a 3D actuated dual-slip
1040–1064. model. IEEE Robotics & Automation Letters 1(2): 1073–1080.
Fevre et al. 17
Manchester IR, Mettin U, Iida F and Tedrake R (2011) Stable Tang JZ, Boudali AM and Manchester IR (2017) Invariant funnels
dynamic walking over uneven terrain. The International Jour- for underactuated dynamic walking robots: New phase vari-
nal of Robotics Research 30(3): 265–279. able and experimental validation. In: IEEE International Con-
Martin AE and Gregg RD (2017) Stable, robust hybrid zero ference on Robotics and Automation, pp. 3497–3504.
dynamics control of powered lower-limb prostheses. IEEE Westervelt ER, Grizzle JW, Chevallereau C, Choi JH and Morris
Transactions on Automatic Control 62(8): 3930–3942. B (2007) Feedback Control of Dynamic Bipedal Robot Loco-
McGeer T (1990a) Passive dynamic walking. The International motion. Boca Raton, FL: CRC press.
Journal of Robotics Research 9(2): 62–82. Westervelt ER, Grizzle JW and Koditschek DE (2003) Hybrid zero
McGeer T (1990b) Passive walking with knees. In: IEEE Interna- dynamics of planar biped walkers. IEEE Transactions on Auto-
tional Conference on Robotics and Automation, pp. 1640– matic Control 48(1): 42–56.
1645. Wisse M (2006) Design and construction of mike; a 2-d autono-
Michel P, Chestnutt J, Kuffner J and Kanade T (2005) Vision- mous biped based on passive dynamic walking. In: Adaptive
guided humanoid footstep planning for dynamic environments. motion of animals and machines. Berlin: Springer, pp. 143–
In: IEEE-RAS International Conference on Humanoid Robots 154.
(Humanoids), pp. 13–18. Yang T, Westervelt ER, Serrani A and Schmiedeler JP (2009) A
Nelson G, Saunders A and Playter R (2019) The PETMAN and framework for the control of stable aperiodic walking in under-
Atlas Robots at Boston Dynamics. Humanoid Robotics: A Ref- actuated planar bipeds. Autonomous Robots 27(3): 227.
erence. Berlin: Springer, pp. 169–186.
Nguyen Q, Agrawal A, Martin W, Geyer H and Sreenath K (2018)
Dynamic bipedal locomotion over stochastic discrete terrain. Appendix A. Index to multimedia extensions
The International Journal of Robotics Research 37(13–14):
1537–1553. Archives of IJRR multimedia extensions published prior to
Nightingale J, Hind R and Goodwine B (2008) Intrinsic vector- 2014 can be found at http://www.ijrr.org, after 2014 all
valued symmetric form for simple mechanical control systems videos are available on the IJRR YouTube channel at http://
in the nonzero velocity setting. In: IEEE International Confer- www.youtube.com/user/ijrrmultimedia
ence on Robotics and Automation, pp. 2435–2440.
Park HW, Ramezani A and Grizzle JW (2013) A finite-state
machine for accommodating unexpected large ground-height
variations in bipedal robot walking. IEEE Transactions on Table of Multimedia Extensions
Robotics 29(2): 331–345.
Extension Media type Description
Post DC and Schmiedeler JP (2014) Velocity disturbance rejection
for planar bipeds walking with HZD-based control. In: IEEE/ 1 Video Uneven terrain (simulation)
RSJ International Conference on Intelligent Robots and Sys- 2 Video Uneven terrains 1–5 (experiment)
tems, pp. 4882–4887. 3 Video Uneven terrains 6 and 7 (experiment)
Saglam CO and Byl K (2018) Quantifying and optimizing robust- 4 Video Random uneven terrain (experiment)
ness of bipedal walking gaits on rough terrain. In: Robotics
Research. Berlin: Springer, pp. 235–251.
Sharbafi MA and Seyfarth A (2017) Bioinspired Legged Locomo-
tion: Models, Concepts, Control and Applications. London: Appendix B. Zero dynamics coefficients
Butterworth-Heinemann.
Shiriaev A, Freidovich L and Manchester I (2008) Can we make a
robot ballerina perform a pirouette? Orbital stabilization of per- a(u) = B? (q)M(F(u))F0 (u)
iodic motions of underactuated mechanical systems. Annual b(u) = B? (q)½C(F(u), F0 (u))F0 (u) + M(F(u))F00 (u)
Reviews in Control 32(2): 200–211.
Shiriaev A, Perram JW and de Wit CC (2005) Constructive tool g(u) = B? (q)G(F(u))
for orbital stabilization of underactuated nonlinear systems:
F(u = ½f1 (u), . . . , fn (u)T )
Virtual constraints approach. IEEE Transactions on Automatic
Control 50(8): 1164–1176. d d2
F0 (u) = F(u), F00 (u) = F(u)
Shiriaev AS, Freidovich LB and Gusev SV (2010) Transverse lin- du du2
earization for controlled mechanical systems with several pas- ð30Þ
sive degrees of freedom. IEEE Transactions on Automatic
Control 55(4): 893–906.