Team RoboSimian: Semi‐autonomous Mobile Manipulation at the 2015 DARPA Robotics Challenge Finals

This paper discusses hardware and software improvements to the RoboSimian system leading up to and during the 2015 DARPA Robotics Challenge (DRC) Finals. Team RoboSimian achieved a 5th place finish by achieving 7 points in 47:59 min. We present an architecture that was structured to be adaptable at the lowest level and repeatable at the highest level. The low‐level adaptability was achieved by leveraging tactile measurements from force torque sensors in the wrist coupled with whole‐body motion primitives. We use the term “behaviors” to conceptualize this low‐level adaptability. Each behavior is a contact‐triggered state machine that enables execution of short‐order manipulation and mobility tasks autonomously. At a high level, we focused on a teach‐and‐repeat style of development by storing executed behaviors and navigation poses in an object/task frame for recall later. This enabled us to perform tasks with high repeatability on competition day while being robust to task differences from practice to execution.