Uncrewed Systems Technology 044 l Xer Technolgies X12 and X8 l Lidar sensors l Stan UGV l USVs insight l AUVSI Xponential 2022 l Cobra Aero A99H l Accession Class USV l Connectors I Oceanology International 2022

53 Stanley Robotics Stan UGV | Digest detect when the frontmost pliers touch the car’s front tyres. Once contact with both tyres is made (with onboard algorithms checking for false positives) the first set of rearward pliers are deployed, and the platform starts to extend backwards until both rear pliers touch the fronts of the rear wheels. Additional onboard algorithms provide a confidence index to gauge how reliable the SLAM readings are in real time, by comparing incoming readings with what should be expected based on the Stan’s position in the known, embedded map. “We also have ultrasonic sensors about the platform as a further verification that we’re maintaining tight contact between the pliers and the wheels, and not straying either side of the central axis,” Trouble notes. “If touch is confirmed across the first four pliers, the final four can then swing into place simultaneously, and the lifting system can do its job.” GNSS When moving between cars, parking spaces, charging points and so on, for SLAM to work reliably the Stan can rely on GNSS for relatively simple localisation and guidance (compared with SLAM). That said, precision guidance and localisation are critical for tight parking of cars and most other driving tasks, so RTK-GNSS is used to ensure centimetric- accuracy of position in real time, with dual antennas for up-to-the-minute heading calculations. Initial GNSS data is received via Trimble antennas and receivers. Every centimetre of space saved during the Stan’s valet parking is additional value to Stanley’s customers, so every centimetre of GNSS accuracy counts. Also, as position and heading are calculated at the head, a tiny error in either reading could result in a bigger mistake being made at the tail end of the platform. A dedicated GNSS ECU takes the signals coming from the GNSS antennas and receivers, and makes them readable by the main PC. The PC also handles the RTK corrections produced by Stanley’s stationary hardware. “Part of the infrastructure for the Stans includes RTK base stations installed on-site to enable the real-time GNSS correction updates,” Trouble says. “We propagate the corrections over wi-fi, and the robots process the corrections to compute the more precise GNSS coordinates and headings.” This localisation information is combined with the SLAM data as well as the odometry coming in from wheel sensors, all on board the main PC. This is to ensure the Stans can constantly navigate roads and parking spaces closely and safely, both when indoors without GNSS and outdoors where SLAM-navigable landmarks might not be available. Computation The main PC is a rugged, embedded computer built around an Intel Core i7 6700 TE CPU. At the time of writing, no GPU had been installed but Stanley expects to choose and integrate one before next year. “The reason for that is that we actually have six very small HD cameras installed about the head of the robot,” Trouble says. “They’re only used right now for remote supervision over wi-fi, so that if the Stan fails to compute something going on around it, we can view what’s happening and potentially intervene from our offices in Paris. “But we’re using machine learning at the moment to develop computer vision for really detailed real-time perception of its surroundings, through which we’ll run the HD video coming from those cameras, and a good GPU will power it. It’s just that graphics cards are currently very expensive and in short supply, so we’re waiting for that to change.” The network architecture running between the computation units forgoes traditional segregation and layering. The entire software architecture is ROS- based, with the necessary timing and arbitration for SLAM and other job-related decision-making defined through the software design. The ECUs meanwhile largely provide acquisition and translation of data between CAN and Ethernet. Trouble notes however, “While the SLAM intelligence and collision avoidance is processed on board, the actual navigation waypoint and commands for the Stan to carry out jobs or head to its charging station are determined by our fleet management system – which interfaces with the end- user’s computer network – and delivered from that program to the local Stan fleet via our comms infrastructure.” Uncrewed Systems Technology | June/July 2022 The company’s SLAM software module is based on the Cartographer algorithm, adapted for localising when minimal landmarks are available