Guessing human intentions to avoid dangerous situations in caregiving robots

TL;DR

The paper tackles the problem of enabling caregiving robots to infer human intentions to avoid dangerous situations. It advances a simulation-based Artificial Theory of Mind (ATM) framework implementing a like-me policy within the CORTEX cognitive architecture, using an internal physics-based simulator to forecast outcomes and test interventions in real time. Two ATM agents perform intention guessing/enactment and action selection, operating on a shared working memory $ \mathcal{W}$ that fuses symbolic and numeric data. Across Webots simulations, human-in-the-loop tests, and a real-world Shadow robot trial, the approach achieves high recall (no missed dangerous intentions) with competitive accuracy (around 79.64%), and sub-second reaction times, demonstrating real-time, safety-oriented intervention in social robot contexts.

Abstract

For robots to interact socially, they must interpret human intentions and anticipate their potential outcomes accurately. This is particularly important for social robots designed for human care, which may face potentially dangerous situations for people, such as unseen obstacles in their way, that should be avoided. This paper explores the Artificial Theory of Mind (ATM) approach to inferring and interpreting human intentions. We propose an algorithm that detects risky situations for humans, selecting a robot action that removes the danger in real time. We use the simulation-based approach to ATM and adopt the 'like-me' policy to assign intentions and actions to people. Using this strategy, the robot can detect and act with a high rate of success under time-constrained situations. The algorithm has been implemented as part of an existing robotics cognitive architecture and tested in simulation scenarios. Three experiments have been conducted to test the implementation's robustness, precision and real-time response, including a simulated scenario, a human-in-the-loop hybrid configuration and a real-world scenario.

Figures (5)

• Figure 1: Schematic view of the model. The robot assigns two targets to the human, the door and the couch, and computes one trajectory in which the person collides with the ball. To save the situation, the robot imagines several possible actions to check if one removes imminent danger. The smaller frame shows the scene rendered by the simulator.
• Figure 2: The CORTEX architecture.
• Figure 3: Combined view showing the contents of $\mathcal{W}$ (up-left); a graphical representation of $\mathcal{W}$ with two paths going from the person to the couch (up-right); a zenithal view of the scene as rendered by Webots (down-left); and a 3D view of the internal simulator, PyBullet, with simple geometric forms representing the elements in the scene (down-right)
• Figure 4: Human-in-the-loop experiment (top left to bottom right). The upper half of each frame is the zenithal view rendered by the Webots simulator. The red and blue axes mark the centre of the scene. The lower half is the view shown to the subject where the ball is missing. When the robot approaches the obstacle, frames 2-3, the subject turns left, overcoming it and safely reaching the couch.
• Figure 5: Real world experiment (left to right). The upper half of the frame shows the view from the robot's camera. The lower half shows a schematic view of the working memory with the person represented as a yellow circle, the backpack on the floor as a red square and the target chair as a green square. The robot is coloured dark red. The subject walks distractedly towards the chair (frame 1) and reacts when the robot starts moving (frame 2), changing direction and continuing.