In embodied artificial intelligence (AI), evolutionary search enables adaptation in complex and uncertain environments but relies on massive stochastic sampling, which in hardware is typically generated using complementary metal-oxide-semiconductor pseudorandom number generators with significant power and area costs. In this paper, we present a 256-magnetic-tunnel-junction (MTJ)-based probabilistic processor as an efficient physical platform for embodied evolution. Using an Ising-based probabilistic sampler, we experimentally construct highly reconfigurable Gaussian probabilistic bits from networks of MTJ-based probabilistic bits and scale the architecture to 256 parallel generators for evolutionary mutation. We demonstrate evolved agent morphologies comparable to those obtained using an ideal software Gaussian random number generator in a robotic locomotion task. MTJ-based stochasticity further drives on-hardware selection, including a diversity-aware scheme that balances lineage diversity and fitness. Our results establish that MTJ-based probabilistic hardware is a promising low-power stochastic substrate for embodied AI and evolutionary robotics.
Bao et al. (Wed,) studied this question.