(2502.03349) Automatic auto goes out from self-playing

(Submitted on 5 February 2025)

See a PDF on paper title title title autonomy emerged from self-play, by Marco Cusumano-Towner and 11 more authors

See PDF
Html (experiment)

Abstract:Self-playing has the power of breakthroughs with two players and multi-player games. Here we show that playing self is a stunning effective strategy of another domain. We show that strong and naturalistic driving emerges completely from playing self to simulation before since scale – 1.6 ~ billion drives. Gigaflow has enabled it, a placing simulator that can synthesize and train 42 years of subjective driving experience every time an 8-GPU node. The resulting policy reaches the state-of-the-art performance of three independent autonomous autonomous benchmarks. The policy faced the first state of the art when tested on recorded world scenarios, between human data, without seeing human data, without seeing human training in the human training during in training. The policy is realistic if it is examined against human references and reached out before death, averaging 17.5 years of continuous driving between simulation incidents.