What do you think about RL learning approaches? Just wondering if that stuff may work in practice. If I'm not wrong Sergey Levine said, that the only problem is sample efficiency. So you would need to simulate the real world, or let your robot brake 1 million time dishes till it learns.