IEEE Robotics & Automation Magazine - December 2021 - 15

(see [4]) to make it more robust to errors as well as to reduce
the risk of accidents.
To obtain the robot pose, we employed visual markers whose
stance can be reliably obtained via calibrated cameras and standard
toolkits. We distinguish markers as static and dynamic,
with associated coordinate frames Sm
We use the notation Tj
and Dm
i for the transformation between coordinate
frames, denoting the translation and rotation of the coordinate
frame i within the coordinate frame j. We can then obtain
transformations TC
Sm and
ST ,Sm
manually. The dynamic marker is placed on the robot and used
to locate its base coordinate frame with respect to camera C via a
constant transformation TR
Dm set manually.
The camera is placed in front of the operational space of
the robot, ensuring proper coverage of the device as well as
the staircase. Its purpose is to associate the robot and staircase
into a common hierarchy of coordinate frames, enabling us,
at any moment, to retrieve the coordinates of the front and
rear step edges with respect to the robot [recall (2)]. The staircase
is then represented as a series of static transformations
S T ,S
i
where Si
is the coordinate frame associated to a step
edge. The staircase perception task could also be performed
by a robot using onboard sensors, but we refrain from doing
so to keep our setup generic and independent of the robot.
Perception of a staircase before descent is more challenging,
but this problem resides beyond our current scope.
Policy Deployment From Simulation to Reality
At the beginning of the experiment, the robot faces the staircase,
and the controller is activated when the device
approaches front and rear step edges. For the task of staircase
negotiation, it turns out that only a subspace of the entire
action space spanned by the flipper and arm joints is useful.
This is something that we take into account by constraining
the latter being measured
(see Figure 3).
the action space from which movements can be sampled. A
direct consequence of the fact that joints are allowed to move
in a significantly smaller space, this provides sufficient time to
perform a necessary action, and adaptation of low-level controllers
from the simulated platform to the real one and
makes this transfer zero shot. Crucially, constraining the
action space further serves to ensure the safety of the platform
since learning is stochastic and previously unencountered
conditions may lead to accidents. For example, overly raising
the rear flippers while ascending or the front flippers while
descending, combined with acceleration, can produce a tipover.
To prevent such behaviors, the limits of the rear flippers
while ascending and the front flippers while descending were
set as ,[ /, ],40
}} ! rrear
front
ss
2
respectively.
The deployed system architecture is described in Figure 3(b).
The marker monitor detects markers in the camera image.
The environment monitor builds and maintains the geometric
relationships between the staircase and robot within a single
transformation hierarchy and provides the front- and
rearmost step coordinates [recall (2)]. The robot state monitor
provides information about the flippers and arm configuration,
the device's velocity, and inertial measurement unit
data. The safety estimator receives the output of the environment
and robot state monitors and evaluates safety metrics,
such as the COG deviation and angular velocity, and its output
is concatenated with data provided by the environment
monitor. That output is fed to the execution controller, which
decides whether to block motors in the case of upcoming
accidents or halt the system when the experiment terminates.
Otherwise, state output data pass to the policy, which samples
the action vector a, which is sent to the robot actuators.
The latter executes those actions, and the marker monitor
observes changes in the environment. Finally, the output of
the safety estimator is recorded via the logger.
Camera, C
Camera Angle
Marker
Monitor
Static Marker, Sm
Logger
Execution
Controller
Policy
Robot State
Monitor
Stair, S
Dynamic Marker, Dm
Robot, R
Environment
(a)
(b)
Figure 3. (a) The experiment setup and coordinate frame hierarchy. (b) The architecture of the deployed robot system.
DECEMBER 2021 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
15
Robot
Legend
Node
Interaction
Interaction
With Reality
Safety
Estimator
Environment
Monitor

IEEE Robotics & Automation Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2021