Amazon Robotics Challenge (2015-2016)
In 2015, I joined the team shortly before the challenge and designed the very simple, yet effective fingers for our parallel jaw gripper. They included a suction cup and a compliant spring steel “spatula” that could be pre-loaded against a shelf to slide under objects. Our manipulation strategy made use of operations we call “primitives” (e.g. horizontal grasping) which is a parameterized behavior made up of sub-actions. In 2016, with the addition of the stowing task, we needed new primitives to effectively pick both downward into the tote and horizontally into the shelf in greater clutter. Our new design had flippable suction cups and actuated, flat-stowing compliant finger tips as well as sensors for detecting contact with the environment and gripper state. Over the course of the year, I created requirements for the gripper based on the primitives we wanted to implement, designed the electrical and mechanical hardware, then fabricated the gripper using a combination of COTS and custom-made parts.
Over the course of this project, I discovered the true cost of not designing for manufacture. It’s easy to say: well, we’re not going to make that many of them, so it’s ok to use glue in this spot, or make this feature annoying to assemble. Ultimately, that’s just lazy design. When the entire assembly is filled with annoying or time-consuming steps (I’m looking at you, epoxy-textile hinges), the cost and the quality of the mechanism really comes back to bite you. I also learned about how error propagates through transmission ratios, non-backdriveability, bi-stability and loads of other more specific topics. As my first year-long project I was in complete control over, it was a real eye-opener in terms of what I can do well and what I can’t. for example, I learned that my optimism made me make more aggressive timelines than I should have. In larger systems problems especially, budgeting for error in time estimation is key.
Motivated by the desire to bring the challenges of robotic manipulation and perception out into the public, Amazon organized the first Amazon Picking Challenge in Seattle, WA in conjunction with ICRA 2015. The challenge brought together robotics research groups from around the world in a simplified version of the real autonomous warehouse picking task. In the first year, robots were required to move ten target items from a vertical shelf to a target tote in twenty minutes. Many teams picked less than two items and we placed second by picking eight. In 2016, the challenge moved to Leipzig, Germany where they introduced the stowing task - where robots must autonomously re-stock the shelf from a jumbled box. Here, we placed 3rd in the stowing task, and 4th in the picking task. To see more about our system in 2017 click here.