IEEE Robotics & Automation Magazine - December 2021 - 16

Experiments
In this section, we present the experiments we performed with
two platforms, namely, Absolem [11] (https://bluebotics
.com and https://github.com/mariogianni/trav_nav_indigo
_ws) and Jaguar V4 Manipulator (http://jaguar.drrobot.com/
specification_V4Arm.asp). Extensive simulated and real-world
experiments were conducted with varying degrees of difficulty
for the task of staircase negotiation, enabling us to successfully
deal with the associated challenges (see the video at https://
partage.imt.fr/index.php/s/LDyp6QRp4nGGKd2/download).
Description of Platforms and Associated Tasks
In Table 2, we juxtapose the two articulated, tracked robots
for which we trained the various controllers. Their simulated
models matched the geometric and mass characteristics of the
Table 2. The robot characteristics and task variants.
Description
Jaguar
Photo
Absolem
real ones, making use of vendor-provided geometric models.
The two robots possess similar obstacle negotiation properties
thanks to the presence of flippers, with the main difference
being that Absolem can independently move each flipper, in
contrast to Jaguar, where the front and rear flippers are coupled.
To make consistent comparisons, we therefore chose to
couple the Absolem flippers, as well.
It follows that each robot has at most 4 DoF corresponding
to linear velocity, angular velocity, and front and rear flipper
angles, with Jaguar having an additional 2 DoF due to its arm.
For the staircase task, though, we decided that yaw angular
velocity control was unnecessary during traversal, after the
following observations: 1) a robot can easily align itself with
the staircase before traversal and 2) it remains aligned during
traversal thanks to the coupled front and rear flippers' grip on
the steps. Thus, the maximum DoF is three for Absolem and
five for Jaguar. Each task in Table 2 is thus characterized by
the direction of traversal (ascent or descent), the DoF of the
action space, and the type of reward used to guide learning
[default refers to the use of (3), COG refers to the use of (4)
and (5), and angular refers to the use of (4) and (6)].
Mass (kg)
Length (m)
Width (m)
Height (m)
Task
Ascent, 3 DoF, default
Ascent, 3 DoF, COG
41
0.98
0.7
0.4
Task ID
Jag-Asc-3-Def
Descent, 3 DoF, angular Jag-Des-3-Ang
Ascent, 5 DoF, COG
Descent, 5 DoF, angular Jag-Asc-5-Ang
29
1.196
0.61
0.46
Task ID
Abs-Asc-3-Def
Jag-Asc-3-COG Abs-Asc-3-COG
Abs-Des-3-Ang
Jag-Asc-5-COG None
None
Simulation Environment
The simulation environment for training is detailed in Figure 4.
A Gazebo simulator (https://gazebosim.org) is employed using
the contact surface motion model of tracks [20] for both robots.
Adding a new robot has no impact on the overall training time
since each controller can be independently trained in batch,
server-side simulations. The simulation environment further
enables us to vary the staircase size in real-world ranges (see [4]
for details). We note that this is an essential step in successfully
transferring policies trained in simulation to the real world, and
it can be thought of as data augmentation that favors generalization
over multiple obstacle shapes and further deals with the
influence of noise in the state estimation process.
The robot starts off at the usual location at either the
Floor
Finish/Start
Stair
Goal
Spawning
Points
Distance
Allowed
Area
ground floor or first floor, depending on the traversal direction,
aligned with the staircase and has to achieve the goal
position by respecting the preset criteria (both positive and
negative rewards are assigned during traversal). The episode
ends if the robot tips over, exits the training area, reaches the
goal, or exceeds the maximal episode time step length. We
evaluate eight task variants, as listed in Table 2, differing by
traversal direction, DoF variation, applied criterion, and
robotic platform. The Jaguar robot model always starts with a
vertically stretched arm that is fixed in 3-DoF tasks. The flippers
of both platforms are parallel to the floor at the beginning,
and ,[ ,] .
zz ! -
aa 90 90
12
c
c
2
Finally, penalties are
evaluated and assigned only when robots are on the staircase.
Results
Start/Finish
Figure 4. The simulation environment.
16 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2021
Ground Floor
Performance for Ascent Tasks
Figure 5(a) charts the mean, smoothed episode return across
independent learning trials, with minimum-maximum bands,
for the policies Jag-Asc-3-COG and Abs-Asc-3-COG. The

https://bluebotics.com/ https://bluebotics.com/ https://www.github.com/mariogianni/trav_nav_indigo_ws https://www.github.com/mariogianni/trav_nav_indigo_ws http://jaguar.drrobot.com/specification_V4Arm.asp http://jaguar.drrobot.com/specification_V4Arm.asp http://partage.imt.fr/index.php/s/LDyp6QRp4nGGKd2/download http://partage.imt.fr/index.php/s/LDyp6QRp4nGGKd2/download https://www.gazebosim.org

IEEE Robotics & Automation Magazine - December 2021

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