Guessing human intentions to avoid dangerous situations in caregiving robots

The paper explores how social robots can interpret human intentions to anticipate and avoid dangerous situations in caregiving environments, building upon the concept of Artificial Theory of Mind (ATM).

The necessity for this work stems from the requirement that For robots to interact socially, they must interpret human intentions and anticipate their potential outcomes accurately. Specifically, caregiving robots must be able to sense and interpret the ongoing activities of individuals in their environment to anticipate future risk scenarios and adjust their behaviour to mitigate the situation in a socially acceptable manner.

The research adopts a simulation-based approach to ATM. The core hypothesis is that simulation-based internal models can provide a feasible basis for ATM, allowing the robots to model human actions and predict consequences. This methodology is known as the like-me simulation, where intentions are assigned by linking people and objects in their environment.

The proposed solution is implemented within a simplified version of the CORTEX cognitive architecture, which utilizes a distributed working memory (W). This architecture employs two primary agents to manage the process:

1. Intention Guessing Agent (Algorithm 1): This agent accesses the current context (ctx) and iterates through all possible interactions between people and objects. It simulates potential actions by considering the person’s gaze (gazel) to determine if an intention is dangerous. The simulation proceeds by executing a path-planning action, and the resulting flag c signals whether a collision occurred.

2. Action Selection Agent (Algorithm 2): This agent activates when an intention is marked as risky. It iterates through the robot's possible actions (Actions) and simulates both the human’s dangerous intention (intijkl) and the robot’s proposed action (robotmn). If a collision is avoided (if c ), that specific action is selected for execution.

The algorithms were tested across three distinct experimental setups: simulation, human-in-the-loop, and real-world deployment.

1. Simulation Experiments (Webots):

The robot analyzed 180 initializations of the scenario where a person was heading toward a target, passing an obstacle.

• The performance metrics showed that Considering the remaining experiments (167), the proposed intention guessing algorithm provides an accuracy rate of 79.64%.

• Crucially, the system demonstrated high reliability by producing no false negatives, i.e., all the unsafe intentions are correctly identified by the robot and no risky situation is left unattended.

In terms of responsiveness, the mean reaction time was less than 0.75s.

2. Human-in-the-Loop Experiment:

Six subjects were instructed to control a person's trajectory in a simulated environment. The robot’s intervention was designed to address an unseen obstacle (a ball).

• The results showed that all subjects avoided the hidden ball when the robot entered their field of view and was detected, successfully completing the task without collisions.

3. Real-World Scenario:

A group of five subjects were instructed to cross a room, ignoring a backpack positioned in their path. The robot was not briefed on its role.

• In all five trials, the robot detected the subject and moved close to the unseen obstacle to prevent the person from stumbling.

The research confirms that an internal physics-based simulator embedded within a cognitive architecture can be a key mechanism for integrating ATM into robotics control. While challenges remain—such as managing the combinatorial explosion of nested relationships, ensuring robustness in real-time execution, and addressing ethical dilemmas when multiple conflicting intentions exist—the the work presented is a new step towards an ATM that can be effectively run inside a robotics cognitive architecture.

Based on the analysis of the Like-Me Simulation-based Artificial Theory of Mind (ATM) framework presented in this paper, the following targeted enhancements are required to address current computational bottlenecks and limitations in real-world deployment.

The current implementation relies on nested loops across People times Targets times Actions times Gaze (Algorithm 1) and similarly complex simulations for action selection (Algorithm 2). This leads to a combinatorial explosion, limiting scalability.

Improvement: Implement heuristic pruning and hierarchical search strategies to aggressively reduce the search space before initiating the simulation phase.

1. Time-Domain Filtering: Instead of exhaustive searching, apply a predictive temporal filter based on kinematic constraints (e.g., maximum human velocity v max and robot response time t response). Any potential trajectory or action that requires a displacement exceeding the collision avoidance time horizon (t) is immediately pruned.

2. Affordance-Weighted Prior Distribution: Assign initial weights to potential intentions based on the object's affordances (e.g., proximity, visual saliency, semantic meaning) rather than treating all targets equally in the initial loop structure. This biases the search toward high-probability interaction points, reducing unnecessary simulations.

3. Constraint-Based Early Exit: If Algorithm 2 successfully identifies a single action that resolves the risk (c) within a limited set of high-probability intentions, terminate subsequent simulations immediately, avoiding exhaustive checking of remaining low-priority actions.

What the Improved AI System Can Do: The system will maintain real-time performance (<0.75s) even in scenes with high density (many people and objects), significantly increasing its operational scalability without sacrificing detection accuracy.

Current intention modeling relies on geometric inclusion within a predefined frustum and uniform sampling of gaze, which is insufficient for robust, real-world deployment.

The current framework is inherently mono-agent, focusing on resolving a single detected risk (int ijkl). It cannot handle situations where multiple individuals pose competing risks.

The current architecture is static in its parameters (e.g., fixed maximum speed, predefined path constraints).