Intelligence
that will eat
the world.

Embodied Intelligence & World Models

Embodied intelligence is studied through agents that learn predictive models of their environments from interaction. The emphasis is on acquiring compact world models from sensory streams such as vision, action, and proprioception, and using these models for planning and control without dense rewards or environment-specific tuning.

Experiments are conducted primarily in simulated and game-based environments, with an emphasis on understanding which learned representations support generalization and long-horizon reasoning.

Neuromorphic Systems

We are currently experimenting with speeding up existing algorithms and architectures to eventually enable back-propagation-free on-chip training of neuromorphic algorithms.

Experiments include using evolutionary strategies to train Spiking Neural Networks without surrogate back propagation, enabling forward-pass on-chip training on neuromorphic chips.

Our SNN speedup experiments with evolutionary strategies at hyperscale are in pre-print — 2.23× faster wall-clock time relative to naive ES.