A Knowledge Driven Approach to Adaptive Assistance Using Preference Reasoning and Explanation

Socially assistive robots (SARs) can aid humans in a variety of tasks. One of the most compelling assistive tasks is in medication management, where a SAR can instruct, record, and oversee a patient’s medication usage. However, since this is a medical application, it is important that a robot is robust, transparent, and open to user feedback; especially for corrections.

However, SARs, like other social robots, are complex to understand. Robots are built of many parts, with underlying language tools (e.g., for NLP, NLU, or NLG) that are not inherently interpretable. Therefore, SARs cannot effectively communicate and collaborate with humans on tasks without explainability. This is troublesome when the robot fails, or when the assistive application is critical, like healthcare or medical applications. In this paper, we make two distinct contributions towards explainable SARs: (1) we contribute a complex cognitive model for incorporating user feedback, and (2) we show a proof-of-concept on a real-life medication case study. Our approach combines three components:

1. 1.

   Preference Reasoning: The robot considers what it knows or can infer about the user’s preferences, and how these affect possible actions.

2. 2.

   Assistance: The robot interacts with the user and aids them if deemed necessary.

3. 3.

   Explanation: The robot explains its reasoning and validates its recommendation and conclusion with a user.

In this paper, we show the capabilities of a SAR with knowledge-driven adaptive assistance. We start with a detailed overview of the medication sorting task. We present the approach and initial results, and conclude with future work, discussion, and a reiteration of the contributions.

We consider a SAR that provides social assistance in a medication sorting task. In this task, a person organizes a set of medications, vitamins, and other supplements. Each has constraints, provided in the form of a prescription, doctor recommendation, or personal preference. In the example we use in this paper, there is a vitamin to be taken each morning and a medication that needs to be taken prior to any physical activity.

The role of the robot is to observe the person while they are organizing the pills into a sorting grid, a device that has compartments for each day of the week and multiple time of day (see Figure 1). Our example uses a sorting grid for four times in the day: morning, noon, evening, and bedtime.

Consider the scenario in which a person is to take a vitamin each morning and then Levodopa before any physical activity. The robot knows that the person has a physical therapy appointment at 1pm on Wednesday and a dance class at 6pm on Friday. The person has a preference that Levodopa is taken enough time before the activity for it to take effect.

After the person has placed one vitamin in the morning compartment for each day, they begin to figure out where to place some Levodopa. The person hesitates, and the robot interjects with a suggestion:

• Robot:

  Try placing a Levodopa pill in the morning on Wednesday.

• User:

  Why?

• Robot:

  Levodopa needs to be taken before any physical activity, and you have a physical therapy appointment at 1pm on Wednesday. Since you prefer to take it a few hours before activity, you should take it in the morning.

• User:

  Oh, right. Thank you.

Alternatively, consider the case in which the person prefers to take Levodopa closer to when the activity is to occur. In this case, the robot would explain:

• Robot:

  A Levodopa pill needs to be taken before any physical activity, and you have a physical therapy appointment at 1pm on Wednesday. Since you prefer to take it immediately before activity, you should take it in the afternoon.

While previous work demonstrated the robot adapting its assistance based on how much assistance the person needed (?), the current work looks at how the robot can reason about the person’s preferences, thus adapting its assistance and explanation.

For the social robot to provide assistance and explanations to the user, we propose an architecture consisting of Preference Reasoning, Adaptive Assistance, and Explanation components, as shown in Figure 2. The Adaptive Assistance component uses updates in the task and social cues conveyed by the user to generate a plan, which is used in determining what action the robot is to take to assist the user. The plan would be sent to the Explanation component, along with the user preferences from the Preference Reasoning component, to generate an explanation for the robot’s method of assistance. Each component is functional individually, but full integration is a work in progress

(prefers user
    (medicationBeforeActivityBy
        medtype
        distance))
 

(prefers user (sortOrder order))
  

(beforeActivity pill row col activity) <-
    (activityAt activity rowX col)
    (isa pill med)
    (medicationBeforeActivityBy med
        distance)
    (difference rowX row distance)

To demonstrate our components adapting to and explaining with user preferences, we consider the scenario in which the user has already correctly placed all of their Vitamin D and is now working to sort their Levodopa, which is to be taken before activity. The user has two activities planned for the week, Wednesday at 1pm and Friday at 8pm. The user just placed one Levodopa pill in the space for Monday midday.

We assume that the user has a previously stated preference to take Levodopa during time slot prior to an activity. When the user misplaces the Levodopa, the Adaptive Assistance component recognizes that the user needs assistance and generates a plan to move the Monday pill to an earlier time slot:

(planFor state8
  ((preference beforeActivity 1))
  ((removePill Levodopa 3 1)
   (addPill Levodopa 3 0)
   (addPill Levodopa 5 2)))

The first action, (removePill Levodopa 3 1), is used along with an inference of a level of assistance to determine that the robot should provide direct assistance, which has the robot clearly stating what should be done next. In this case, the pill needs to be removed and the robot would say “Try removing a Levodopa from Wednesday”.

Additionally, the Adaptive Assistance component considers alternative plans, a counterfactual reasoning over different preferences. The alternative plans do not affect how the robot assists but would be sent to the Explanation component. An alternative plan for taking Levodopa at the same time as the activity is shown below. The plan indicates that the Monday pill is in the correct location and the only action remaining is to place the Friday pill:

(alternativePlanFor state8
  ((preference beforeActivity 0))
  ((addPill Levodopa 5 3)))

Finally, the user can inquire about the robot’s actions. For example, the user can ask why the robot said to “Try removing a Levodopa from Wednesday.” This question is parsed into an intermediate representation: (onDate Levodopa Wednesday), which is passed to the Explanation Synthesizer along with the associated preference; the user prefers to take the medication before an activity. The following is a trace of the reasoning of the Explanation Synthesizer

[(IsA Levodopa pill), ’Given’]
[(AtLocation pill cabinet), ’ConceptNet’]
...
[(IsA Wednesday weekday), ’ConceptNet’]
[(IsA Wednesday day), ’ConceptNet’]
...
[(prefers user (before pill activity)),
    ’Given preference’]
[(IsA appt activity), ’Given knowledge’]
[(atTime appt ’1pm’), ’calendar’]
[(onDay appt Wednesday), ’calendar’]
[(atTime appt afternoon), ’Rule fired’]

The justification is that (onDay pill Wednesday) (beforeTime pill afternoon). And the final explanation reads in a series of symbolic triples: (prefers user (before pill activity)) (IsA user activity) (atTime appt ’1pm’) (onDay appt Wednesday) (IsA ’1pm’ afternoon).

The work we describe here sets the foundation for a whole line of work in designing social robots to adapt to users, adhere to user preferences, and provide explanations. The most immediate next step is to build upon the proof of concept we have demonstrated here by integrating the individual parts and evaluating the system on more complex scenarios.

Once we have a fully integrated system, the critical next step is to incorporate feedback from the user. One of the greatest advantages of taking a knowledge-driven approach is that the entire system is inspectable, which will facilitate integrating the user feedback. The user may provide feedback after the robot provides an explanation to the user. An explanation synthesizer extracts key terms from the feedback, interprets the feedback, and determines which component(s) need adjustment. The explanation synthesizer also validates its conclusion by verifying with the user.

In this ongoing work, if the explanation synthesizer identifies that the feedback is related to the user preferences, the Preference Reasoner will construct a new case that is used by AToM to update the model of the user. Even a single piece of feedback from the user can be sufficient for AToM to learn the user’s preferences because analogical learning, which is used by AToM, is data efficient and capable of learning from only a few examples (?; ?).

Consideration of user preferences is an important aspect of human-robot interaction, as it allows the robot to modify its behaviors according to its understanding of the user. Hiatt, Harrison, and Trafton (?) found that people prefer collaborating with robots that adapt their behaviors in this way. However, most approaches to recognizing users’ preferences use statistical techniques like reinforcement learning (?) and Markov Decision Processes (?) to predict preferences. This means that they require large amounts of training data, are not responsive to user feedback, and are not explainable. That is, once trained, such systems predict preferences based on their built-up statistical models; a user cannot state a preference directly or inspect why the robot predicts a particular preference. By incorporating stated user preferences, and moving toward learning preferences by analogy, we attempt to avoid these pitfalls.

There are many forms of adaptive assistance in robotics. One approach is shared autonomy, in which the system infers human intentions and adapts how much assistance in provided in controlling a robot (?; ?). This work is focused on assisting people in physically controlling robots, whereas we are working towards an autonomous robot that provides social assistance.

Other work has looked at adapting a robot’s behavior based on user preferences. For example, a recursive neural network was used to learn weights pertaining to user preferences, which influence the plans for a robot (?). While the preferences did affect the plans used by a robot, the plans are used to improve the robot’s navigation. Thus, the user preferences do not relate to assistance provided to the user.

A model of Theory of Mind (ToM) has been proposed to adapt the assistance provided by a social robot (?). A stochastic model is used to infer what action a person could be executing. Based on this estimate, they generate a plan to determine with which action the robot should help. While they use ToM to estimate a user’s intent (via a set of possible actions), they do not represent a user’s preference for how the task should be completed.

One way to understand complex decision making systems is with interpretable or explainable parts. Explanations can describe proxy methods (?; ?; ?), representations (?; ?), or be inherently explanation-producing (?). In the context of human-robot interaction, explanations can help to communicate and build trust (?), justify the robot’s actions (?) or motions (?), or describe unreasonable perceptions (?). But most of these explanations are generated after-the-fact and cannot be used to improve the completion of tasks moving forward.

We propose to use explanations as feedback to augment assistive robots. This has been explored for agents playing games, especially when to provide explanations (?). This approach builds on Rainbow, a self-adaptive system that can correct itself and reuse the same baseline framework (?). To our knowledge, this is the first work to propose a knowledge-driven architecture that could use explanations to improve robotic reasoning and inference.

In this paper, we motivate a knowledge-driven architecture for adaptive assistance. We demonstrate the functionality of the components of this architecture in a task for socially assistive robots (SARs). In future work, we will expand the architecture to incorporate and process feedback and learn user preferences. This paper opens a new area of research in adaptable and interpretable SARs.