A quadruped learns to walk in one hour
DayDreamer applied a world-model algorithm directly to physical robots, with no simulator and no resets: a quadruped learned to roll off its back, stand up and walk from scratch in one hour.
Why it matters
Learning on hardware was considered impractical because of the cost of attempts; one hour of real time from an empty policy to walking changed what counts as practical. There is no simulator in this loop at all, and so no transfer gap to cross.
Philipp Wu, Alejandro Escontrela, Danijar Hafner, Ken Goldberg and Pieter Abbeel took the Dreamer algorithm and ran it on four different platforms. A quadruped learned to roll off its back, stand up and walk from scratch in one hour, without resets, and within a further ten minutes adapted to withstand perturbations. Two robot arms learned to pick and place multiple objects directly from camera images and sparse rewards. A wheeled robot learned to navigate to a goal position from images alone, resolving the ambiguity about its own orientation by itself. The same hyperparameters were used across all four platforms, which was one of the paper’s claims.