To the construction site of the autonomous cleaning robot developed by PFN and Kashima.Recognize the surrounding environment with deep learning

AI cleaning robot "raccoon" and autonomous movement system "iNoh"

AI cleaning robot "raccoon" (raccoon)

First, about the AI ​​cleaning robot "raccoon". "Raccoon" is a robot that cleans dust and dirt on the concrete floor while moving autonomously with a minimum of 3 touch instructions from the operation screen of the main unit.

The size of the main body is 1,205 mm in length, 776 mm in width, and 816 mm in height. It weighs about 80 kg and is lighter than it looks. The base was a ready-made vacuum cleaner with wheels, a motor, and a control unit. There is a bucket on the bottom for collecting garbage, and the capacity is 37 liters. It fits within the 900 mm frame, which is the size of a general door frame, and is said to be the standard size for a cleaning machine.

"Raccoon" front side rear side. The gray part is the underside of the bucket racoon.Modified ready-made vacuum cleaner

The battery uses two lithium-ion batteries with a rated capacity of 9.3Ah (413Wh) in parallel. The continuous operation time is about 120 minutes assuming 400W. Approximately 100 minutes at full operation of 500W. Charging time is about 1 hour (time to change the remaining battery power from 10% to 80%). It can detect steps of 10 mm or more and can overcome steps of up to 12 mm by self-propelled.

The main sensors are Livox's 3D LiDAR (laser sensor) in front, a fisheye lens camera, Intel Realsense to see your feet on both sides, and Hokuyo's 2D LiDAR in front.

In addition to this, the bumper sensor that detects physical contact is front and back, and the ultrasonic sensor for detecting obstacles is 3 in front and 1 in back. And five infrared sensors for step detection are attached.

The IMU (Inertial Sensor) uses the built-in IMU of 3D LiDAR. The operation panel is located on the top of the main unit and can be moved manually using the joystick as well as the start setting.

Various sensors are lined up.The center is empty because it is under test, so Livox's 3D LiDAR control panel has a joystick that can be operated manually.

Processing for autonomous movement and deep learning is performed by a dust-compatible industrial PC mounted inside the robot. SLAM (processing that simultaneously creates a map and estimates self-position) is performed by the CPU, and deep learning processing is performed by the GPU.

We plan to optimize it in the future, but the current specifications are Core i9-9900 (8 cores / 16 threads, 3.1 GHz, cache 16 MB), GPU is GeForce GTX 1660 SUPER. Logs are collected via LTE.

The cleaning method is a sweeper method, and dust is taken in with the brushes on the bottom and sides. The cleaning width is 620 mm. It has two cleaning modes. "Random cleaning mode" that autonomously cleans while searching for a cleanable area by itself without a map of the site or instructions from workers, and a cleaning area from the construction drawing that is linked after automatically creating a map of the cleanable area. There are two "area cleaning modes" that can be specified.

It can be operated without a map, but at the time of operation, a pre-map is first created using a robot and aligned with the 2D drawing.

When moving the robot, it first indicates which floor it is on, and the robot reads the pre-map accordingly. During operation, it operates while updating the pre-map.

If there is no prior map, clean it while making a map. Basically, it aims to be able to clean by giving only an instruction "which floor you are on" and turning on the switch.

The cleaning area is said to be 200 to 400 square meters in 2 hours. According to the results of trial introduction of "raccoon" at multiple sites in the metropolitan area, it was possible to clean an area of ​​about 500 square meters with 100 minutes of continuous operation. As a result, both companies have confirmed the practicality of the autonomous mobile system "iNoh".