also factored in vehicle dynamics, but WeRide gradually decided it still wasn’t accurate enough. “We realised one day that what we really needed were the intentions of other vehicles – do they want to bypass us, cut in or just follow the road?” Liu recounts. “That was when we introduced machine learning into our prediction model: to learn, based on preceding movements, what each vehicle’s driver was intending. Maybe six months later, we found we could very consistently predict high-level intentions, but what about the low-level – meaning is that vehicle to your forward-right going to try and cut in harshly or just gently? “If it’s the former, we need to brake much harder and earlier than for a gentle cut. And we can’t just hard-brake every time to cover both kinds of cut-in, because sooner or later that will either injure our passengers or cause collisions with the vehicles behind us.” Hence, based on map location and vehicle positions, the prediction model also calculates 12 possible trajectories (including velocities) of moving obstacles around the ego, each with an assigned score (aggregating its probability, as well as accounting for how dangerous its consequences could be). Naturally, each trajectory’s score updates to become a more accurate reflection of reality as more real-world traffic data is digested by the cloud IDE. “Training the prediction model took millions of hours of well-distributed driving data, both from our uncrewed and driven vehicles, and we constructed a neural network with a transformertype architecture, and designed multidimensional optimised objectives so that the models could come to reasonable decisions in all space distributions,” Liu says. “Neural networks are interesting, but we’re convinced that data itself is the more important part of machine learning. Cleaner, better-distributed data is much more helpful to an intelligent model than having the best kind of neural network, and more and more neural networks used today are open-source, transparent solutions that anyone can get hold of. Your data is really what sets you apart.” Compute systems The self-driving algorithms run on a centralised computer (the aforementioned MU), with dedicated lower-level computers managing the cabin systems, comms, battery, brakes, steering motors, drive motor and inverter. The current MU design is referred to internally as HPC (high-performance computer) 2.0, and it was based on an x86 architecture with integrated CPUs and GPUs. A backup computer to the MU also provides redundancy, taking over from the MU to run the autonomy WeRide Robobus | Dossier microDRIVE LP hargravetechnologies.com The Next Generation of Motor Controller 60A Continuous Current 60V Operating Voltage IP67 Protection DroneCAN & ARINC 825 +Dshot, PWM NDAA Compliant hrgve.tech/ust-58 True to Size Download the Datasheet
RkJQdWJsaXNoZXIy MjI2Mzk4