Page 1 of 14
Transactions on Machine Learning and Artificial Intelligence - Vol. 10, No. 1
Publication Date: February, 25, 2022
DOI:10.14738/tmlai.101.11616. Niijima, S., Sasaki, Y., & Mizoguchi, H. (2022). Real-Robot Friendly Passing Motion Planner for Autonomous Navigation in Crowds.
Transactions on Machine Learning and Artificial Intelligence, 10(1). 27-40.
Services for Science and Education – United Kingdom
Real-Robot Friendly Passing Motion Planner for Autonomous
Navigation in Crowds
Shun Niijima
Department of Mechanical Engineering
Tokyo University of Science
2641 Yamazaki, Noda-Shi, Chiba, 278-8510, Japan
Yoko Sasaki
National Institute of Advanced Industrial Science and Technology,
2-3-26 Aomi, Koto-ku, Tokyo 135-0064, Japan
Hroshi Mizoguchi
Department of Mechanical Engineering
Tokyo University of Science
2641 Yamazaki, Noda-Shi, Chiba, 278-8510, Japan
ABSTRACT
This study proposes a real-robot friendly passing motion planner to be used in
crowds. The proposed method learns to pass pedestrians with smooth acceleration
and deceleration by using passing motion learning. A key feature of the proposed
method that it is trained on a simple crowd simulation with both of dynamic and
stationary pedestrians. The learned passing behavior can be used directly in an
autonomous navigation. Evaluations using the crowd simulations indicate that the
proposed method outperforms the existing ones in terms of success rate, arrival
time, and keeping a certain distance from the pedestrians. The proposed navigation
framework is implemented on a mobile robot and demonstrated its successful
navigation between pedestrians in a science museum.
Keywords: passing motion plan, dynamic environment, mobile robot, deep reinforcement
learning
INTRODUCTION
Mobile robots are required to operate in close proximity to
people, such as in shopping malls or exhibition halls. To
move efficiently, many mobile robots use pre-generated map
information of the environment to calculate a path, avoid
obstacles, and move to its destination.
As shown in Figure 1, there are many pedestrians in a real
environment, and a mobile robot needs to detect and avoid
them with its on-board sensors to run safely. Motion
planners usually make use of established parameterized
approaches, such as the dynamic window approach (DWA) Figure 1. Mobile robot in crowds.
Page 2 of 14
28
Transactions on Machine Learning and Artificial Intelligence (TMLAI) Vol 10, Issue 1, February - 2022
Services for Science and Education – United Kingdom
or elastic bands [1], [2], which assume that the current position of the obstacle is static.
However, in reality the obstacles can be dynamic as well. When using these methods, the robot
assumes that the pedestrians are stationary, hence, it may get too close to them or even collide
with them.
Velocity obstacle (VO) [3] was proposed to consider dynamic obstacles. VO is a planner that
generates avoidance maneuvers by selecting the robot velocities outside the collision cone,
which consists of velocities that would result in collision with obstacles moving at given
velocities sometime in the future. The authors of [4] proposed reciprocal velocity obstacle
(RVO), which is an extension of VO in a multi-agent environment. In RVO, each agent moves by
considering the behavior of other agents to achieve a mutual collision avoidance. Moreover, the
authors of [5] used a variant of RVO, namely, the optimal reciprocal collision avoidance (ORCA)
to pro-actively avoid pedestrians. These behavioral policies enable the robot to move efficiently
even among dynamic pedestrians. However, while these policies are sufficient in an artificial
environment where all agents perform uniform movements, they are less optimal in the real
world because pedestrians do not always move as expected. In addition, the motion model of
the robot is not considered, thus, the robot may accelerate and turn rapidly and is forced to run
dangerously.
There has been numerous research on methods for motion planning in dynamic environments.
Several methods have improved performance by incorporating pedestrian motion models into
navigation [6-8]. In [9], a socially aware recurrent neural network model is used to predict
trajectories and set feasible sample waypoints as the optimal path. In [10], a Gaussian process
(GP) is used to learn the motion patterns of pedestrians, and a path planning is performed by
incorporating probabilities into the cost function of random rapid tree search. Although these
methods are effective to ensure that the robot keeps an appropriate distance from the
pedestrians, the freezing robot problem (FRP) [11] may occur when faced with dynamic
pedestrians. For example, because the pedestrian prediction model contains uncertainty, the
robot may not be able to find a safe path through the crowd.
This study proposes a learning method for the passing motion based on deep reinforcement
learning (DRL). An approach of end-to-end motion planner based on DRL is very effective in
difficult to model environments, such as crowds with many pedestrians. The proposed method
trains a passing motion using a simple pedestrian passing simulation. The learned passing
motion planner can be directly integrated into the navigation of a mobile robot, and
experimental results have been obtained on a real robot.
RELATED WORK
With the rapid developments in the machine learning field and DRL, researchers have recently
started using neural networks for robot navigation in dynamic environments [12]. These end- to-end navigation learning methods depend on the environment of the training data, hence, it
is difficult to use them outside the trained environment. Therefore, they are combined with
path planning methods, such as A*, to learn only the motion generation [13,14].
The method described in [15] includes a distributed multi-agent collision avoidance algorithm
based on deep reinforcement learning. This research was later extended by modifying the
reward function to ensure that the agents adhere to social norms [16]. Because these models
Page 3 of 14
29
Niijima, S., Sasaki, Y., & Mizoguchi, H. (2022). Real-Robot Friendly Passing Motion Planner for Autonomous Navigation in Crowds. Transactions on
Machine Learning and Artificial Intelligence, 10(1). 27-40.
URL: http://dx.doi.org/10.14738/tmlai.101.11616.
focus on moving pedestrians, they suffer from stoppages and collisions when faced with
unrecognized pedestrians or stationary obstacles. In a real environment, these problems need
to be addressed because some people stand still for various reasons, for example, to look at
exhibits. Therefore, the aim of this research is to add stationary pedestrians to the learning
process to ensure that the system learns to behave effectively in real environments.
In many cases, the network output is a rough discretized value, such as linear velocity or
angular direction at a certain time. If the obtained behaviors are directly applied to the velocity
commands of the robot, sudden robot velocity and pose changes may occur. Such sudden
changes in speed makes pedestrians uncomfortable, therefore are not suitable for robots that
operate among them. The proposed passing motion learning system learns to accelerate and
decelerate smoothly from the current speed as the robot navigates. In addition, it uses a grid
map of the robot's surroundings as its current state and outputs speed control commands
directly from its input. This eliminates the need for dedicated nodes to calculate velocity
commands and can be easily replaced with existing motion plans.
The main contribution of this study is that it proposes a transit motion learning method that
can be directly integrated into the autonomous navigation of real robots. The main features of
the proposed method are following: 1) the separation of the passage motion from the path
planning makes the learning motion effective in various environments. 2) the method learns
not only from moving pedestrians but also from stationary pedestrians. 3) the passage motion
planning can learn smooth acceleration and deceleration motions to calculate the velocity
command outputs. In the demonstration experiment at the exhibition, the effectiveness of the
proposed method is confirmed by implementing it on a real robot.
PASSING MOTION LEARNING FOR MOBILE ROBOT NAVIGATION
Simulation environment with pedestrians passing by
The learning of passing motion is carried out using a simulation that represents the crowds
with pedestrians and robot. To simulate the passing motion of a robot and a pedestrian, it is
required that pedestrian and robot move purposefully in a limited space. The simulation model
represents pedestrians with a simple circle model. To produce a real-like crowd, the proposed
simulation includes stationary pedestrians as well.
Figure 2 (a) shows the representation of a pedestrian model in simulation, where the position
of the pedestrian is represented as a green circle with a radius �! . The pedestrian circle is
inflated by the radius of the robot �", which denotes the collision area. The predicted range of
pedestrian movement is drawn in the direction of the pedestrian velocity as a cone with a
central angle of 30 degrees. To learn moving without disturbing pedestrians, the simulation
includes the predicted area. The predicted area representation of the 30 degrees cone is
constructed by referring to the forward angle of the pedestrian field of view in the pedestrian
modeling described in [17]. The size of prediction area is scaled using the pedestrian velocity.
The color of the predictive area is represented by a grayscale with a 2 second range at every 0.5
second.
The simulation considers two environments according to the placement of pedestrians. One is
an unobstructed space (figure 2 (b)), while the other contains randomly placed stationary
pedestrians (figure 2 (c)). Walking pedestrians move toward the opposite side from where they