Play Data

Robot data recorded while an operator freely manipulates a scene with no specific task in mind.

Play data is a collection approach introduced by Corey Lynch and colleagues in the 2019 paper Learning Latent Plans from Play (CoRL 2019): an operator teleoperates a robot around a scene full of objects, opening drawers, pushing sliders, and grabbing blocks out of curiosity, with no task specified in advance. Its advantage is cost: there's no need to segment by task, label anything, or reset the scene between attempts. The paper reports that, for the same amount of collection time, this approach covers roughly 4 times the range of interactions that task-by-task demonstrations do. The tradeoff is that there are no task labels, so this data is usually paired with hindsight relabeling (treating the endpoint of a segment as its goal) to train a goal-conditioned policy, or given a language description after the fact. The CALVIN benchmark and MimicPlay both build on this idea.

ExampleIn the Play-LMP paper, an operator teleoperates a simulated robot through VR, freely manipulating a tabletop scene with drawers, sliding doors, and blocks; the resulting continuous data, with no segmentation or task labels, is used to train a policy that can complete various tasks given a goal image.

Also called

Free Play Data

Related

Goal-conditioned Policy · Hindsight Relabeling · Teleoperation · Demonstration Data · CALVIN Benchmark · MimicPlay

Sources

Learning Latent Plans from Play (arXiv:1903.01973)