How to build a mind: Exploring insect-inspired AI for autonomous robots

Quick Overview

The presentation advocates for a brain-inspired approach to AI autonomy, contrasting it with current, highly resource-intensive methods like SLAM and deep reinforcement learning, by highlighting biological systems (insects, sea squirts, birds) that solve complex navigation problems with far greater efficiency and robustness, suggesting that future AI should mimic this sparse, efficient processing.

Key Points: Humanity has long dreamed of creating artificial intelligences that coexist with us, but current AI approaches like SLAM and deep reinforcement learning are computationally intensive, requiring massive resources (e.g., the Apollo program cost $288 billion in inflation-adjusted dollars). Biological systems offer inspiration for efficient autonomy; for example, fruit fly brains solve navigation problems using only about 100,000 neurons in a cubic millimeter, contrasting sharply with contemporary AI resource demands. Tunicates (sea squirts) solve movement first, then perception, digesting their own brain once they settle down to free up resources, illustrating efficiency through selective resource management. Kestrels stabilize their visual perception by locking their heads in three dimensions while hovering, solving the inverse kinematics problem simply to keep their head stable while looking for prey. The speaker proposes that AI should move away from massive data and compute reliance, focusing instead on sparse, biologically inspired circuits that solve the specific problems of spatial intelligence and navigation robustly, as demonstrated by the fly brain's navigation circuit. The company Optran offers a vision-only alternative to SLAM, aiming to solve localization, mapping, and navigation in challenging environments (scene change, featureless spaces, reflections, glass, dynamic lighting) while drastically cutting compute and sensor costs. The speaker concludes by stating that the goal is not Artificial General Intelligence (AGI) but robust, simple, and efficient algorithms inspired by nature to solve real-world autonomy challenges.