The Badb Loop

In adversarial environments, the speed of your learning loop determines whether you keep the initiative - or cede it to your adversary. Within the Badb project, Rigr AI has a particular responsibility: detecting and tracking vehicles, especially military vehicles.

One obvious way to do this is to collect and annotate real-world data: video from operational environments and train models on the resulting datasets. This approach has many benefits - increasing domain knowledge, capturing real equipment, tactics, and their evolution under the pressure of an intensely adversarial environment.

This approach also has a significant problem: if we rely only on real-world data, we are constrained by whatever we can collect. Even assuming that security classifications are not an obstacle, this creates a fundamental limitation: our adversary becomes a limiting factor on data collection and model training cadence.

When our adversary introduces a new vehicle, a modification of it, or a new tactic, we see it only when we encounter it. It takes time to collect enough examples to adapt a model. Bluntly, we might find that we can only ship an effective updated model after giving the enemy plenty of time to employ their clever new vehicle/mod/tactic. It is extremely undesirable to be confined to a reactive role - ceding the initiative to the adversary.

Our learning loop must allow us to adapt and improve faster than our adversary. To guarantee a fast-learning loop, we cannot rely on real-world data alone. We need a machine that allows us to generate data, train models, evaluate them rigorously, and deploy them continuously.

So, we are building a software platform, Badb Loop. The loop combines synthetic and real-world data. We use simulation (of environments and scenarios informed by real-world data) to generate large volumes of controlled, richly annotated data and train new models. Evaluation spans both simulated and real data. Where performance is insufficient, we iterate generating targeted data to address specific weaknesses. Models eventually graduate to deployment at the edge. Real-world performance can then feed back into the system, informing the next round of simulation, training, and evaluation.

Simulation gives us control over crucial variables - vehicles, structures, terrain, weather, and time of day. We can explore edge cases that might be rare or dangerous to encounter in the real world, generate accurate labels, and scale to the limits of our computing infrastructure.

We want to be clear that simulation on its own is not enough. In a truly adversarial environment, adaptation will be rapid and will draw on the full spectrum of possibilities, and so the scenario library will always need updates. Over time, this library will come to represent a rich bank of challenges for UAVs and UGVs, allowing us to deliver robust AI for perception and control of UxVs - and to complete future adaptation cycles with higher quality and faster iteration times.The purpose of Badb Loop is simple: to ensure that our rate of learning is not constrained by our adversary, and, better still, that we outpace their ability to adapt.