RoboticsDexterous learningStory 03
Minutes of play seed hours of robot practice
What changedDSP behavior-clones minutes of human play, lets the robot collect hours of its own interaction data, then uses the learned model for planning or distillation. In real experiments it beat hindsight behavior cloning by 69.6 percent in success rate and matched reinforcement learning with 490 times less interaction data.
The useful part
Why it matters
In real experiments it beat hindsight behavior cloning by 69.6 percent in success rate and matched reinforcement learning with 490 times less interaction data.
Keep in mind
Good to know
The 490-fold comparison depends on the authors' task and interaction accounting.
Evidence