Getting a Grip on Robotic Data Collection

One of the biggest challenges in robotics right now isn't the hardware. It's data. While many data collection methods are effective, handheld data collection can create a diverse dataset of environments, conditions and strategies for completing manipulation tasks. Teaching a robot to turn a bolt or pick a single nail from a pile requires thousands of reliable demonstrations. However, if you build a gripper for a robot and then design a separate handheld interface to collect data, it can leave you with demos that don't transfer well. The proprioceptive feedback the human feels through the interface doesn't reflect what the robot will actually sense. The Koala platform: co-designed the handheld grippers and robot grippers around the same linkage mechanism, same degrees of freedom, and same force transmission. The human feels through the linkages what the robot will feel through its actuators. Learn more about the our gripper designs, the data we collect, how to train policies with that data, and where we are taking the grippers next: https://rai-inst.com/resources/blog/h... 00:00:57 Robotic Manipulation Challenges 00:01:54 Compliance vs. Rigid Manipulation 00:03:02 The Need for Massive, Diverse Data 00:03:52 Overview of Data Collection Methods 00:06:23 Scaling Data Collection with Handheld Devices 00:07:22 Why Position Alone Isn't Enough 00:09:03 The Benefits of Co-Design 00:09:20 Introducing the Koala Platform 00:12:00 Versatility of a Single Gripper Across Tasks 00:13:53 Proprioceptive Feedback Through Linkages 00:14:37 Curated demos vs. Unstructured "In the Wild" 00:17:08 What's Next and Closing Thoughts