BC-Z

A 2021 Google dataset of real-robot demonstrations across 100 tasks, collected to study zero-shot task generalization.

BC-Z is a project from Google and X, Alphabet's “moonshot factory” division, presented at the robot-learning conference CoRL 2021; the name stands for “behavioral cloning + zero-shot.” Using 12 robots and 7 operators, the team collected demonstrations through VR teleoperation combined with “shared autonomy” — the policy acts on its own, and a human takes over just before it would fail, a method called HG-DAgger — gathering 25,877 demonstrations totaling 125 hours across 100 tasks, plus 18,726 videos of humans performing the same tasks. The policy receives a language or human-video task embedding through FiLM layers (a way to inject a conditioning signal into a neural network) and reached 44% average success on 24 held-out tasks it was never trained on. It was an early demonstration that multi-task data plus language conditioning produces zero-shot generalization; the dataset was later folded into the Open X-Embodiment collection.

ExampleAt test time, the policy is given instructions involving object combinations it never saw paired together during training — for instance, placing an object into a container it was never matched with — and must complete the task with zero additional demonstrations. Across 24 such novel tasks, it reached 44% average success.

Also called

BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning, BC-Z Robot Dataset

Related

Open X-Embodiment · RT-1 · Human-Gated DAgger · Language-conditioned Policy · Zero-shot · Feature-wise Linear Modulation

Sources

BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning (CoRL 2021, PMLR)
BC-Z 项目主页 (Chinese)

As of

2022-02