98 UVIO | Swiss-Mile UGV “That’s where reinforcement learning [RL] had a huge impact, enabling the robot to figure out good locomotion by itself, once we’d defined the parameters and constraints on what it couldn’t do, as well as the optimisation objectives. It marked a breakthrough in our autonomous locomotion and performance.” With Swiss-Mile’s own autonomy stack and intelligence installed, the UGV becomes a unique autonomous system in its own right. To mark the intelligence and capabilities of its robotic dog apart from the B2-W, Swiss-Mile has named the new system Milo. Training Milo We previously saw RL being used to great effect by Spanish firm Keybotic in training its Keyper quadruped UGV (Issue 52), and Wellhausen puts forward two reasons why it is proving revolutionary in engineering legged robots. “When RL policies are trained correctly they can handle uncertainty a lot better than other forms of motion control can achieve,” he says. “RL neural networks internalise uncertainty about their environments while learning, making them very robust. Basically, by training Milo in environments where it was subjected to external disturbances, it learned how to bypass, work around or ignore those disturbances.” Conversely, model-based approaches require that the engineer model such disturbances, and then install some means for the UGV to detect and classify those disturbances, overall entailing far more engineering labour than RL does. “Second, with RL, you don’t have to design the motion manually. Trying to manually specify when the robot should step, and when it shouldn’t, brings forth a huge combinatorial problem, because this robot has four legs, and each leg requires a decision to be made based on both directional commands and terrain properties,” Wellhausen explains. This requires immense heuristics, including matters such as step detection (for intelligent reaction in case a leg gets stuck, for instance), so being able to learn the right reactions itself via RL makes developing the UGV far easier. Having developed its algorithms in this way on a modified research quadruped, the company was swift to install and adapt them into the new B2-W platform to create Milo, with the first prototype entering in-the-loop simulations after roughly five days of work, and hardware locomotion tests after three weeks. “Of course, we had a six-year r&d background at ETH Zurich, which helped us know going in what was necessary to make this work in hardware, including defining constraints and rewards for the machine-learning network to achieve good motion,” Wellhausen says. “With the B2-W’s morphology being fairly similar to the modified research quadruped, it was fairly easy to switch it to a new platform. If we had to move to another new one, we’d probably be pretty fast still.” Hard hats on Among Swiss-Mile’s target use-cases, construction stands out as an industry that uncrewed systems have had trouble breaking into, owing to a variety of regulatory and technological hurdles. Despite these, construction companies are inviting Swiss-Mile to demonstrate Milo’s capabilities at project locations. One key application on site is safety monitoring, where a high-resolution camera is installed on Milo to take videos and photography to verify that everything is up to code, such as the accessibility of emergency exits, the correct placement of heavy equipment and the installation of guardrails. Another is progress monitoring, which would see Milo running controlled routes with cameras (or Lidar for high-fidelity digital twins) to identify work progress since the last survey, and also to compare the new 3D model with the customer’s digital model to see if construction is proceeding correctly. “Rework is one of the construction industry’s biggest sources of cost. One estimate puts up to 30% of construction costs being just rework costs, because incorrectly built sections get disassembled, materials go to waste and significant work has to be done all over again. So, if Milo can be used to catch errors earlier, when they’re still small, it works out much cheaper for August/September 2024 | Uncrewed Systems Technology Payloads are mission- and customer-specific, while the robot chassis is Unitree’s B2-W and the software is 100% Swiss-Mile’s (Image courtesy of Swiss-Mile)
RkJQdWJsaXNoZXIy MjI2Mzk4