FRESHR-GSI: A Generalized Safety Model and Evaluation Framework for Mobile Robots in Multi-Human Environments

TL;DR

The work targets safety in mobile robot operation amid multiple humans by introducing the Generalized Safety Index ($GSI$), a directional, multi-human safety metric derived from distance $d_{h,r}$, relative velocity $v_{h,r}$, and bearing $\theta_{h,r}$. Implemented in the RGB-D based framework FRESHR, $GSI$ supports robot-centered and external-observer viewpoints and aggregates per-human risk via a LogSumExp with temperature $\tau$, while a tunable exponent $\rho$ shapes conservativeness. Validation across real robots, simulations, and crowd datasets shows that $GSI$ captures nuanced risk in crowded scenes and improves over averaging-based safety scales. The framework enables safety-aware motion planning and multi-robot coordination in human-rich environments, with practical tunability through $\rho$ and $\tau$.

Abstract

Human safety is critical in applications involving close human-robot interactions (HRI) and is a key aspect of physical compatibility between humans and robots. While measures of human safety in HRI exist, these mainly target industrial settings involving robotic manipulators. Less attention has been paid to settings where mobile robots and humans share the space. This paper introduces a new robot-centered directional framework of human safety. It is particularly useful for evaluating mobile robots as they operate in environments populated by multiple humans. The framework integrates several key metrics, such as each human's relative distance, speed, and orientation. The core novelty lies in the framework's flexibility to accommodate different application requirements while allowing for both the robot-centered and external observer points of view. We instantiate the framework by using RGB-D based vision integrated with a deep learning-based human detection pipeline to yield a generalized safety index (GSI) that instantaneously assesses human safety. We evaluate GSI's capability of producing appropriate, robust, and fine-grained safety measures in real-world experimental scenarios and compare its performance with extant safety models.

Figures (8)

• Figure 1: Overview of the proposed vision-based safety evaluation framework (FRESHR). A continuous sequence of RGB images is supplied to Yolov7 for human and robot detection and human pose estimation. These detections and their confidence values will be integrated with depth information to calculate key metrics such as the relative distance and velocity between humans and robots. Finally, a normalized safety value will be provided to evaluate the multi-human safety of mobile robots.
• Figure 2: FRESHR aligns its safety scale with the known and empirically determined proximity ranges in human-robot interaction spaces. GSI takes a value of 1 indicating safe (green) in the public space, (0 - 1] (amber) in the personal and social spaces, and 0 (red) in the intimate space.
• Figure 3: GSI can be fitted to various applications, robot platform properties, and subjective safety perceptions of humans through parameter $\rho > 0$. For instance, $\rho=1$ is set for assessing safety, $\rho>1$ for more cautious robot control, and $\rho<1$ for a more closer interaction with human who are already comfortable.
• Figure 4: ($a$) An example setting with three humans in the vicinity of the mobile robot $r$. FRESHR yields a directional safety value for each human. In this example, $GSI_{h_1} = 0.7$ with $\theta_{h_1,r} = 290^\circ$, $GSI_{h_2} = 0.9$ with $\theta_{h_2,r} = 345^\circ$, and $GSI_{h_3} = 0.4$ with $\theta_{h_3,r} = 30^\circ$, each of which is calculated using Eq. \ref{['eqn:gsi_single_thetar']}. ($b$) Impact of the hyperparameter $\tau$ on the collective $GSI$ (polar plot in (a)) obtained using Eq. \ref{['eqn:gsi_overall']}. The individual GSIs used are those from ($a$).
• Figure 5: ($a$) Human moves along three trajectories in the shared space: straight toward the task robot (setting 1), perpendicular to the task robot (setting 2), and diagonal (setting 3). ($b$) FRESHR-based estimation of the distances (left) and relative velocities (Right) from two different viewpoints.