RoboNet

A 2019 dataset of multi-robot interaction video, about 15 million frames across 7 kinds of robots.

RoboNet is an early large-scale multi-robot dataset released in 2019 by researchers at Berkeley, Stanford, Penn, and CMU, including Sergey Levine and Chelsea Finn. It pools about 15 million frames of video and corresponding actions from 7 robot platforms, ranging from the industrial Kuka arm to the low-cost WidowX, as well as the Sawyer, Franka, Baxter, Fetch, and Google's R3. The data was collected autonomously by having robots execute random actions, requiring almost no human demonstration and carrying no task labels; it's mainly used to train action-conditioned video-prediction models for visual foresight planning. The paper found that pretraining on RoboNet and then transferring to a new robot outperformed training from scratch on that robot with 4 to 20 times as much data. It was an early attempt at a cross-embodiment dataset.

ExampleA researcher collects only 300–400 random trajectories on a new robot arm, fine-tunes a video-prediction model pretrained on RoboNet, and then uses it to plan pushing actions on objects.

Also called

RoboNet: Large-Scale Multi-Robot Learning

Related

Cross-Embodiment Data · Visual Foresight · Video Prediction Model · Autonomous Data Collection · Open X-Embodiment · Pre-training

Sources

RoboNet: Large-Scale Multi-Robot Learning (arXiv:1910.11215)
RoboNet 项目主页 (Chinese)

As of

2020-01