The main input to the AV system as a whole consists of an array of sensors typically including cameras, LIDARs, RADARs, and many others dedicated towards building a comprehensive understanding of the vehicle’s immediate surroundings. The various streams of raw sensor data are fused together using combinations of classical and AI methods. This allows the AV system to classify objects of interest like road boundaries, lane markers, vehicles, pedestrians, traffic signs, and more as well as tracking previously observed objects and localizing the vehicle itself within a local or global map. Large challenges facing perception systems are dealing with sensor noise and occlusion.
Once the AV system has perceived the objects and determined their current states within its environment, it must next predict all relevant future states. For example, an AV system needs to predict whether the vehicle ahead of it will slow down or continue its current speed which will influence its own behavior. In AV 1.0, this has often involved simplified kinematic models, predefined behavioral rule sets, or basic machine learning models. While effective in many scenarios, these traditional methods can struggle to capture the full spectrum of human variability, nuance, and decision-making complexity, particularly in dynamic, ambiguous situations. This is precisely where Inverted AI offers a transformative enhancement.
The AV systems, equipped with information about the states of objects in its environment, must now generate a trajectory to achieve its navigational goal (e.g. a given GPS coordinate) given some set of constraints (e.g. speed limits, general safety, efficiency). In typical AV 1.0 systems, this involves a complex combination of tasks such as global pathfinding, local maneuver generation (e.g., lane changes, merges), and collision avoidance. This is where the Inverted AI tools will have the most benefit by generating trajectories that account for full coverage of highly dynamic scenarios that are difficult to observe during design and testing of AV 1.0 planners.
The control module takes the planned trajectory (e.g., desired steering angle, acceleration/braking profile) and translates it into precise, real-time commands for the vehicle's actuators (e.g., steering wheel, brakes, throttle). This ensures the vehicle accurately executes the trajectory determined by the planner.
At Inverted AI, our predictive Imagine the Road Ahead (ITRA) Generative AI model has been developed to generate human-like behaviors in simulated environments. Recently, our team adapted this model to achieve the prediction and planning components of an AV 1.0 system and verified its functionality onboard a real, on-the-road platform. Request the application note Implementing a <u>State-of-the-Art Autonomous Vehicle Planner</u> with Inverted AI to learn more about implementing ITRA into your AV system today; the most advanced model for predicting the behavior of all humans on and near the road and planning safe, smooth, and realistic trajectories in complex urban or high-speed highway settings with ease all in one package.