We are developing a new class of trainable inverse solvers designed to help AI systems perceive and interact with the physical world.
Many real-world tasks – perception, control, and decision-making – require solving inverse problems: inferring hidden causes from observations, such as estimating an object pose from a 2D image or determining the actions needed to reach a desired state. Current AI approaches often rely on iterative optimization or large learned models, which can be computationally expensive and difficult to orchestrate reliably across tasks.
Our approach moves beyond purely loss-driven learning. Instead, we use geometric principles to make convergence toward the correct solution structurally enforced rather than probabilistic. This allows the forward model matching procedure to converge rapidly and opens a path toward single-step, noise-robust solutions to inverse problems. More broadly, we are exploring a path toward deterministic, noise-resilient inference mechanisms for physical AI systems.
Our first prototype demonstrates this approach by recovering CAD parameters directly from LiDAR measurements for industrial quality inspection.
From CAD to Quality –
Without Manual Tuning
LiDAR-Fit turns raw LiDAR scans into CAD-accurate deviation analysis – robust to noise, misalignment, and production variability.
Built on principled inverse solvers, the system converges to the correct alignment without false trapping. Residual errors stem solely from measurement noise and model mismatch – not from optimization artifacts.