IEEE Robotics & Automation Magazine - March 2021 - 56

dynamics-based MPC is a standalone lateral controller,
where the output is a control quantity of the steering angle.
In such a case, a longitudinal PID controller that outputs
the speed control quantity will be required. And the outputs of these two controllers constitute u(t).

immediately taken over by a human driver in the case of
any failure for autonomous navigation. We expect less
human intervention during our goods transportation
tasks. The performance was evaluated according to the
number of occurrences of human interventions.

Evaluation
This section describes the real tasks of contactless goods
transportation using our vehicle during the COVID-19 pandemic in China. From 2 February to 27 May 2020, we
deployed 25 vehicles in three different cities (Zibo, Shandong; Suzhou, Jiangsu; and Shenzhen, Guangdong) in
China. Our current server can handle 200 vehicles simultaneously. The total running distance of each vehicle reached
2,500 km. The details of the transportation tasks are summarized in Table 1. Selected
demonstration photos of
Our DBW chassis can be
the tasks are displayed
in Figure 7.
divided into four parts:
Note that, in the case
of failure during autono1) motion control, 2)
mous navigation, such as
unavoidable accidents
electronic accessories
and system errors, we
built a parallel driving syscontrol, 3) basic sensors,
tem to back up our vehicle control. The parallel
and 4) system control.
driving system is a remote
control system based on
4G/5G technology. When
using 4G with a good signal, the latency is usually between 30
and 60 ms. We set the system to automatically adjust the bit
rate to ensure that the vehicle is not out of line. When using
5G, the latency can be fewer than 20 ms. If the vehicle is out of
line, it will stop immediately.
We tested the function with several vehicles on several
real road environments, and the experimental results
indicate that the function works well. Vehicle control is

Lessons Learned and Conclusions
For object detection, we found that, in practice, real-time performance deserves much more attention than accuracy. The
perception algorithms should be efficient since they lie in the
front end of the whole autonomous navigation system. Therefore, we replaced the dense convolution layers with spatially
sparse convolution in our 3D object detection module. As a
result, the inference time was boosted from approximately
250 to 56 ms.
For the point-cloud mapping, we found that our system
was capable of dealing with the rapid motion of the vehicle
and short-term point-cloud occlusions. Since most of the
lidars are mounted parallel to the ground and the vehicles
always move along the ground, the typical ring structure
of the point clouds makes the system difficult to observe
in terms of translational movements vertical to the
ground plane. Drift in this direction is inevitable during
long-term operations. In practice, we learned that the
GPS localization results signaled the potential loop, leading to a consistent larger-region map for the lidar-based
localization. In very crowded dynamic traffic environments, the system could be degraded by disturbances
from moving objects [10]. To tackle this issue, we used
semantic segmentation to remove movable objects to
obtain clear point-cloud data.
For the planning part, we found that four-layer hierarchical planning is necessary and effective. The working
environments of our vehicles are complex in terms of
road structures, traffic conditions, driving etiquette, and
so on. Layer-wise planning makes the system extensible
for multiple environments and various requirements.
Furthermore, it is essential to attach importance to the

Table 1. Details of the contactless goods transportation tasks during the COVID-19 pandemic in China.
City

Task

Time
Distance Duration

Payload

Environments

Characteristics

Zibo, Shandong

Vegetable delivery

9.6 km

75 min

600 kg

Urban main road

Light traffic, heavy payload,
and slow speed

Vegetable delivery

5.4 km

50 min

960 kg

Urban main road
and residential area

Light traffic, heavy payload,
and slow speed

Meal delivery

1.2 km

30 min

80 kg (100 boxes
of meals)

Urban main road

Medium traffic, left turn,
and U-turn

Road disinfection

0.6 km

10 min

Residential area

Narrow road, crowded,
obstacle, and slow speed

Meal delivery

1.6 km

20 min

64 kg (80 boxes
of meals)

Urban main road
and residential area

Heavy traffic, narrow road,
and U-turn

Meal delivery

4.0 km

40 min

96 kg (120 boxes
of meals)

Urban main road
and residential area

Heavy traffic, narrow road,
and U-turn

Suzhou, Jiangsu

Shenzhen,
Guangdong

IEEE ROBOTICS & AUTOMATION MAGAZINE

MARCH 2021

IEEE Robotics & Automation Magazine - March 2021

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