3-D Perception, Mapping, and Localization
October 25, 2013, Robo Business, Santa Clara, CA—Michael Bruenig from the Commonwealth Scientific and Industrial Research Organization in Australia presented their findings on vision and mapping. The main idea is to free the robots from their tethers.
The current big driver for autonomy is in mining. The shippers want to have robot workers doing the dangerous work without humans. The context for developments shows the possibilities. Sensors in cars manage short-and long-range optical and RF detectors, GPS, and inertial guidance, all leading to a full 3-D map of the local environment.
The sensor arrays are able to handle dynamic environments and situations like pedestrians, emergency traffic, etc., but the implementers need to reduce costs and complexity to make the technology more widespread. The ability to localize the vehicle and maintain some reasonable track also requires some means to develop a map with minimal prior knowledge of the area.
The tasks for localization and mapping need the ability to scale to any size environment. The key is in containing the false representations. With LIDAR, the vehicle can only get a 3-D perspective, but adding mapping enables a global context for locality and mapping.
The structures are fixed in mines, so developing a map only needs the LIDAR to detect objects in the space. In one trial, they mapped 17 km of mines in about 2 hours. For a large cave, they mapped the space of and matched the detailed existing maps that took years to complete. The mapping of 9 km took about 10 hours.
The effort led to changes in the sensor package. The first version used a laptop computer and an oscillating LIDAR on a small cart to discern the full 3-D object field. Part of the reason for the oscillatory motion is to improve the vertical image resolution. The desire to lighten the load and allow the mapper to operate in much smaller spaces led to a hand-held unit. The smaller unit also needs to have inertial guidance capabilities, since the sensor array is hand held.
Normal body motions and spring mounting of the sensors provide the vertical motion needed for full 3-D imaging. The nice outcome is that the hand-held array doesn’t require any calibration. The operator just walks with the equipment and takes readings. The commercial LIDAR scanner has 6 mm accuracy.
The possibility for personal localization and mapping just needs some manufacturer to convert their prototype into a final product. Other obvious apps include manufacturing and forestry.
The key to the effectiveness is a continuous-time algorithm that can maximize the information from a single sensor. Adding the mapping capabilities to a robot will enable more autonomous functions and allow the robots to work in unrestricted spaces.
