[Paper Review] Planning for Proactive Assistance in Environments with Partial Observability

Anagha Kulkarni, Siddharth Srivastava+1 more|arXiv (Cornell University)|May 2, 2021

AI in Service Interactions4 citations

TL;DR

This paper proposes a Monte Carlo Tree Search (MCTS)-based framework for proactive AI assistance in partially observable environments, where the agent reduces human task cost by strategically modulating human belief states through communicative actions. It demonstrates via user studies that legible, belief-modulating actions significantly improve human awareness of assistance and reduce perceived workload and planning effort.

ABSTRACT

This paper addresses the problem of synthesizing the behavior of an AI agent that provides proactive task assistance to a human in settings like factory floors where they may coexist in a common environment. Unlike in the case of requested assistance, the human may not be expecting proactive assistance and hence it is crucial for the agent to ensure that the human is aware of how the assistance affects her task. This becomes harder when there is a possibility that the human may neither have full knowledge of the AI agent's capabilities nor have full observability of its activities. Therefore, our \ extit{proactive assistant} is guided by the following three principles: \ extbf{(1)} its activity decreases the human's cost towards her goal; \ extbf{(2)} the human is able to recognize the potential reduction in her cost; \ extbf{(3)} its activity optimizes the human's overall cost (time/resources) of achieving her goal. Through empirical evaluation and user studies, we demonstrate the usefulness of our approach.

Motivation & Objective

• To address the challenge of proactive assistance in environments where humans have limited observability of AI actions and may not expect help.
• To ensure that proactive assistance is both effective in reducing human task cost and recognizable by the human to prevent confusion or resistance.
• To optimize the overall cost (time/resources) incurred by the human by balancing assistance benefits with cognitive overhead from processing the agent's behavior.
• To develop a framework that enables the AI agent to reason over human belief states and control observability to communicate cost-reduction benefits effectively.
• To validate through empirical evaluation and user studies that legible, belief-modulating actions enhance human awareness and reduce perceived task load.

Proposed method

• The framework uses Monte Carlo Tree Search (MCTS) to sample and evaluate partial joint plans between the human and AI agent.
• It constructs a utility tree that evaluates assistive behaviors based on their impact on human cost, awareness, and overall effort.
• The agent synthesizes communicative behaviors—such as displaying items or announcing cleared rooms—to modulate the human’s belief state and increase recognition of assistance.
• The system controls human observability by selectively revealing or hiding information (e.g., showing medkit in wagon, indicating empty rooms), reducing cognitive load.
• It incorporates epistemic actions (e.g., speech acts) and ontic actions with epistemic effects to ensure the human can infer the assistance’s benefit.
• The method balances three principles: reducing human cost, enabling awareness of cost reduction, and minimizing overall human effort from processing the agent’s behavior.

Experimental results

Research questions

• RQ1 Can a proactive AI assistant effectively reduce human task cost in partially observable environments where the human may not expect assistance?
• RQ2 To what extent do legible, belief-modulating actions improve human awareness of proactive assistance?
• RQ3 How does the inclusion of communicative behaviors affect the perceived workload and planning effort of the human?
• RQ4 Can the agent’s behavior be synthesized to optimize both task cost reduction and cognitive overhead for the human?
• RQ5 What is the impact of strategic information disclosure (e.g., showing items, indicating cleared rooms) on human task performance and perception?

Key findings

• In user study 1, 25 out of 31 participants recognized the robot’s assistive behavior when a legible action (displaying wagon contents) was included, compared to only 6 out of 31 in the baseline without it.
• In user study 2, 24 out of 27 participants recognized the robot’s assistance with a legible action (announcing cleared rooms), versus 6 out of 27 in the baseline.
• The average workload score for the proactive assistant (PA) condition was 5.96 (on a 7-point scale), significantly higher than the baseline’s 3.22 (p < 0.0000001, Cohen’s d = 1.89).
• The average effort to process the robot’s behavior was 6.06 for PA versus 3.45 for baseline, with a p-value < 0.05 and effect size of 1.62.
• In study 2, PA reduced perceived workload to 5.55 (vs. 2.74 for baseline), with a p-value < 0.05 and effect size of 1.95.
• The average processing effort was 5.85 for PA versus 3.55 for baseline, with a p-value < 0.05 and effect size of 1.67, confirming reduced cognitive overhead with legible actions.