Issue 40 Unmanned Systems Technology October/November 2021 ANYbotics ANYmal D l AI systems focus l Aquatic Drones Phoenix 5 l Space vehicles insight l Sky Eye Rapier X-25 l FlyingBasket FB3 l GCS focus l AUVSI Xponential 2021

28 Dossier | ANYbotics ANYmal D many sub-processes,” Fankhauser says. “Most attempts at legged robots use a model-based approach, with a state estimator algorithm running at around 200 Hz to fuse all the velocity and position information, and so on. “That certainly works, but we found through our r&d at ETH Zurich that with a learning-based approach, we can achieve far more robust locomotion and hence autonomy, as well as doing things that engineers using a model-based approach couldn’t even begin to form rules for.” ANYbotics’ locomotion stack has therefore been developed by having the robot ‘learn’ how to walk from hundreds of thousands of physics-based simulations and units of input data. A ‘reward’ function is used to ensure it can identify the best rules and policies for achieving stable balance and motion. Then, through a proprietary process devised over its years of r&d, the company can transfer those rules and policies from the simulation engine into the robots’ CPUs. Getting such transfers right is typically extremely difficult, not so much from having to copy code but more because of the inaccuracies of simulated physics, motor dynamics and various other qualities. Fankhauser remarks that the ability to know which parts of the simulations to carry over has “blown our past efforts at locomotion out of the water”. He adds, “It’s a robust and fast set of algorithms, but most important it’s computationally lightweight. That means plenty of CPU space is left over to manage the other subsystems.” The motor controllers and actuators have been custom-designed and optimised for torque and position measurement (as well as tight integration) over many years. They communicate using EtherCAT with the main computer, as this bus system provides real-time capability with a high data throughput. Each SEA drive contains a permanent magnet motor, an inverter, a gearbox and a set of sensors for torque and position, all monitored by its own low-level CPU. Dr Mauerer adds, “We’re working closely with Maxon for our next- generation custom motor drive, including the inverters and gearboxes, since they are quite specialised in tailoring drives for high-performance robotic applications. “It’s an impedance-controllable drive, if you will, where you can actively control the mechanical impedance. That’s possible precisely because we have torque sensors in the drives.” Navigation and localisation As well as keeping itself balanced from one step to the next, the ANYmal D must be able to localise itself within a given map, for autonomous navigation as well as accurate reporting of survey results along its mission waypoints. Naturally that requires first having a map embedded in its 240 Gbyte SSD drive. There are two ways of doing that. The first is that if the customer already has a complete 3D model of their facilities, or at least all the routes they want the ANYmal D to survey, they can upload that to the UGV. The second approach is for one of the customer’s engineers to remotely pilot an ANYmal (using a ruggedised handheld GCS) on an initial mapping route through their facility. As the engineer stands a few paces away from the robot, steering it through corridors and equipment, the various cameras October/November 2021 | Unmanned Systems Technology The company designed and carried out numerous testing profiles to accurately gauge the shock and vibration limits of the robot An ANYdrive motor undergoing submergence testing