Aligning Learning with Communication in Shared Autonomy

TL;DR

This work addresses aligning learning with communication in shared autonomy for assistive robot arms. It combines experimental and theoretical analysis to show that explicitly communicating the robot's learned intent alters human input patterns, and that the robot can leverage this to better align with the user's goal. The authors formulate a Bayesian, maximum-entropy framework where the human is modeled as a Boltzmann rational agent and the robot updates a posterior over goals while blending its own actions with human input; communication enters the human cost via a term involving the robot's belief. They demonstrate, through online and in-person studies, that a joint learning-communication approach reduces user effort and improves subjective and objective performance compared to baselines that isolate learning or communication. The results suggest that explicit communication of learned intent, when coupled with learning that accounts for user responses to communication, can significantly improve the efficiency and intuitiveness of assistive robotic systems.

Abstract

Assistive robot arms can help humans by partially automating their desired tasks. Consider an adult with motor impairments controlling an assistive robot arm to eat dinner. The robot can reduce the number of human inputs -- and how precise those inputs need to be -- by recognizing what the human wants (e.g., a fork) and assisting for that task (e.g., moving towards the fork). Prior research has largely focused on learning the human's task and providing meaningful assistance. But as the robot learns and assists, we also need to ensure that the human understands the robot's intent (e.g., does the human know the robot is reaching for a fork?). In this paper, we study the effects of communicating learned assistance from the robot back to the human operator. We do not focus on the specific interfaces used for communication. Instead, we develop experimental and theoretical models of a) how communication changes the way humans interact with assistive robot arms, and b) how robots can harness these changes to better align with the human's intent. We first conduct online and in-person user studies where participants operate robots that provide partial assistance, and we measure how the human's inputs change with and without communication. With communication, we find that humans are more likely to intervene when the robot incorrectly predicts their intent, and more likely to release control when the robot correctly understands their task. We then use these findings to modify an established robot learning algorithm so that the robot can correctly interpret the human's inputs when communication is present. Our results from a second in-person user study suggest that this combination of communication and learning outperforms assistive systems that isolate either learning or communication.

Figures (4)

• Figure 1: Human sharing control with an assistive robot arm. (Top) The robot tries to infer the correct task from the human's joystick inputs. (Middle) We show that --- when the robot communicates what it has inferred --- the way humans provide inputs changes. (Bottom) If robots are aware of these changes, they can more accurately infer the human's goal.
• Figure 2: Example settings and results from our user studies in Section \ref{['sec:user-study-one']}. Here we explored how communicating the robot's inferred distribution over a discrete set of tasks affected the human's inputs during shared autonomy. In all conditions, the robot used the same learning algorithm. (Left) Results from the online survey with and without a communication interface. Humans were more likely to release control to an assistive robot that conveyed its learned distribution over the tasks ($t(24) = 4.271, p < 0.005$). (Right) Corresponding results from our in-person study. Here humans required fewer inputs to guide the robot to their goal when the robot communicated its learning ($t(29) = 2.986, p < 0.005$). Overall, these results suggest that humans are more willing to yield control to a communicative system. An asterisk (*) denotes statistical significance.
• Figure 3: Tasks and user inputs from the user study in Section \ref{['sec:user-study']}. (Left) The items the human led the robot to interact within each task. (Right) The magnitude of the human's inputs over time averaged across all users. These results show that users completed the tasks more quickly with Ours, and overall needed fewer inputs to convey their intended goals to the robot.
• Figure 4: Objective and subjective results from the user study in Section \ref{['sec:user-study']}. (Left) Total user inputs for Seasoning, Drink, and Utensil tasks. To count the number of inputs, the robot measured whether the human had pressed the joystick every $0.02$ seconds. Across each task, users provided fewer inputs and relied on the robot's assistance more when using Ours ($p < 0.001$, $p < 0.001$, $p < 0.001$). These results support H3: Users spent less effort when using Ours. (Right) Subjective results for the three baselines. Across the four Likert-Scale items, users preferred Our method: they felt that they could easily control the system ($p < 0.001$), the robot provided effective assistance ($p < 0.005$), the robot better predicted their goal ($p < 0.001$), and the robot adapted more quickly to their actions ($p < 0.001$).