The progress of deep reinforcement learning in tackling complex problems such as robotics and gaming is impressive, but the challenges of gathering highly varied, large and complex datasets limit its potential.