The increasing availability of low-cost, compliant and human-friendly manipulators allows robots, such as Rethink Robotics' Baxter, to be placed in close proximity to human workers. Unlike traditional automation systems, which needed to be kept in cages, these compliant robots can share a common workspace with human workers. A clear benefit of this close proximity is the opportunity for cooperation between a human worker and an assistive robot.
In order for a robot to be effective in an assistive role, it is important that the human is able to easily and quickly understand the robot's intentions by observing its actions. Ideally, this understanding will come in an intuitive manner, similar to how humans are innately able to communicate with one another non-verbally when working in close quarters. Legible motion plans are an important part of making the robot understandable by human co-workers intuitively. In this context, the legibility of a motion corresponds to whether human subjects can realize early on which is the actual target of the moving arm out of many possible choices.
The goal of this work is to identify the key features of robotic motion for manipulators that contribute to their legibility. The motions are executed by two seven degrees-of-freedom manipulation arms that are mounted on a Baxter robot. The arms move towards grasping multiple targets, which are positioned linearly in front of the robot. As the manipulator moves, human subjects observe the robot and report their belief regarding the intended target of the arm. An appropriate experimental setup was developed in order to collect these responses, so as to minimize the cognitive load of the human subjects and achieve good accuracy.
The five different types of trajectories that were considered during this experimental study cover a variety of discriminant legibility features. Some of the features correspond to arm policies, such as the shortest path in the configuration space, and other correspond to hand, i.e., end-effector, policies, such as the orientation of the end-effector relative to the target. The experimental results show that the legibility of different trajectories is indeed different and consistent across different targets. Motions which allow the end-effector to point towards the intended target and move along a straight line in the workspace result in enhanced legibility.
The longer term objective of identifying these legibility characteristics is the design of motion planners that incorporate these features into the planning process so as to automatically generate legible motion. The hope is that co-robots, which can generate legible motion plans, can more effectively collaborate with humans.
In order to evaluate key features of legible paths for manipulators, this experimental study considers a variety of path types for a dual arm manipulator, i.e., a Baxter robot by Rethink Robotics. There are four arm policies considered in this study:
1. shortest path in configuration space (i.e., minimizing change in joint angles),
2. overhead motion frequently appearing in "pick and place" paths,
3. shortest, straight-line path for the end-effector in workspace, and
4. "curved" path for the end-effector in the workspace to exaggerate intent.
Likewise, this study considers two potential hand policies:
1. hand goes immediately to final joint position (e.g., overhead grasp) and stays there for the duration of the motion, and
2. hand points toward the goal in the workspace at all times. The pointing feature of these paths can be seen as a symbol generating anticipation of the motion.
Given these two factors, there were five different types of trajectories generated:
1. "Shortest" path: This is the shortest path in the configuration space computed on an asymptotically near-optimal version of a probabilistic roadmap in the Open Motion Planning Library. This class makes use of arm policy 1 and immediately provides a path for the hand as well.
2. "Overhead Down" path: Similar to paths employed for pick-and-place tasks by axter robots in industrial settings, where the end-effector moves in a position over the target and points downwards throughout the motion. This class makes use of arm policy 2 and hand policy 1.
3. "Straight" path: The robot moves its end eector along a linear path from the initial position to the target object while the end effector points towards the target. This class uses arm policy 3 and hand policy 2.
4. "Straight Down" path: The robot moves its end effector along a linear path from the initial position to the target object while the end effector remains in a vertical orientation pointing down. This class makes use of arm policy 3 with hand policy 1.
5. "Curved" path: The robot moves its end eector along an exaggerated curved path while pointing at the target. This class is inspired by ideas in previous work towards generating legible paths. This class makes use of arm policy 4 combined with hand policy 2.