Generative ai for embedded applications

Generative AI for better perception and planning

by industry veterans

CEO Tyler Marchand and CTO Francois “Frank” Belletti, PhD

Motivation

Embodied AI presents unique challenges from limited compute and energy to very expensive failure modes once deployed in the real world. We are here to make that “rubber hits the road” moment more seamless and build systems that automatically identify and fix their shortcomings.

From model efficiency to synthetic data generation based on standard procedural principles and diffusion models, creative solutions to old problems enable hitherto unseen adaptivity and robustness in efficient models.

Fixing well known PITFALLS with original solutions

Models don’t know what they don’t know. And the typical scaling solutions for better generalization are not always compatible with embodied deployment.

This is where we come in. From improved efficiency in edge models to automated model verification and improvement processes, we use old and new techniques in tandem to make AI world in the real world, within realistic deployment energy and compute budgets.

EXpanding tried-and-true solutions

From experience on Youtube neural recommendations with fast yet long-range dependent models, to Kubric and MoveNet, and more modern improvements that made state-of-the-art diffusion models like Veo 3 and Genie 3 considerably faster, we’ve learned a thing or two along the way when it comes to making models faster, more robust, and adaptive.

In parallel to Levenlight, we are developing some simple tutorials to help newcomers understand the techniques and approaches that make a drastic difference when it comes to making AI work for real in actual applications involving unknown unknowns.

MAKING WORLD MODELS TRAIN AI THAT ACTUALLY WORKS

World models are particularly exciting as universal simulators which can be used to trained agents to achieve complex tasks as in this amazing paper.

However, the sim-to-real gap is more complex to grasp and the issue of reward hacking more prevalent than ever. This is where more traditional simulation techniques combined with agentic procedural generation can make a world of difference.