Robots just like babies, do not know much about the world around them. They need to learn crowd preferences if they wish to move around, without annoying or harming people. This is a difficult problem because the criterion that defines a good trajectory can vary with tasks and environments.
PlanIt is a crowdsourcing system to collect large amount of preference data by showing videos to non-expert users. PlanIt learns a cost function from the preference data and uses it to generate desirable robot trajectories. Since the users are not experts, the feedback is weak and noisy but our extensive experiments show that this feedback helps generate preferred trajectories. We also visualize the learned cost functions as heatmaps and confirm that the preferences learned make intuitive sense.
Also special thanks to
for their help in setting up and testing the system.