Fairness risk and its privacy-enabled solution in AI-driven robotic applications

TL;DR

The paper investigates fairness risk in AI-driven robotic systems and proposes a utility-aware fairness framework that accounts for user utility and inherent data randomness. It establishes formal local $g$-fairness and global $g$-fairness metrics and reveals a principled link to differential privacy, showing that smaller privacy budgets bound the fairness metrics, e.g., $\bar{L}(P,g) \le L(P,g) \le \varepsilon_A + \log\left(1 + \frac{L_A \cdot diam(\mathcal{A}) + \delta_A \gamma}{\tau}\right)$. A privacy-based remedy is proposed, where DP on sensitive attributes via privacy filters enforces fairness guarantees, validated through a vision–language model–driven robot navigation case study on the S3DIS dataset. The results indicate that carefully chosen privacy budgets can jointly satisfy privacy and fairness targets, offering a practical pathway for trustworthy and ethically aligned autonomous robotics.

Abstract

Complex decision-making by autonomous machines and algorithms could underpin the foundations of future society. Generative AI is emerging as a powerful engine for such transitions. However, we show that Generative AI-driven developments pose a critical pitfall: fairness concerns. In robotic applications, although intuitions about fairness are common, a precise and implementable definition that captures user utility and inherent data randomness is missing. Here we provide a utility-aware fairness metric for robotic decision making and analyze fairness jointly with user-data privacy, deriving conditions under which privacy budgets govern fairness metrics. This yields a unified framework that formalizes and quantifies fairness and its interplay with privacy, which is tested in a robot navigation task. In view of the fact that under legal requirements, most robotic systems will enforce user privacy, the approach shows surprisingly that such privacy budgets can be jointly used to meet fairness targets. Addressing fairness concerns in the creative combined consideration of privacy is a step towards ethical use of AI and strengthens trust in autonomous robots deployed in everyday environments.

Figures (5)

• Figure 1: System architecture of the proposed robot navigation framework. A point cloud map is used to generate top-view and topological representations. Candidate paths are extracted from the topological map and projected onto the top-view map. A vision–language model then selects an optimal route from these candidates, which the robot follows to complete the navigation task.
• Figure 2: Unfairness results in LLM-based robot navigation. Unfairness is observed in LLM-based robot navigation even if the document types are different, as the GPT models consistently select HR_2, leading to an unfair workload.
• Figure 3: System Architecture of privacy filters. Raw features $X$ and the sensitive attribute $A$ from the agent are privatized by separate filters to $\tilde{X}$ and $\tilde{A}$, satisfying $(\varepsilon_X,\delta_X)$-DP and $(\varepsilon_A,\delta_A)$-DP, respectively. Because the attributes $X$ and $A$ are privatized before being transmitted to the robotic system, the VLM-driven robotic system generates its response $U$ according to the distribution $P(U \mid \tilde{X}, \tilde{A})$.
• Figure 4: Fairness-Privacy results in LLM-based robot navigation. In this task, $L(P,g)=0$ indicates a fair workload across document types, while $\bar{L}(P,g)=0$ denotes fairness in the average workload aggregated over the three document types. In this experiment, we set $\delta_A = 0$. As shown, the fairness metric increases with $\varepsilon_A$, indicating that the privacy parameter directly influences fairness; stronger privacy guarantees therefore promote fairer outcomes.
• Figure 5: System architecture of the proposed fairness-aware robot navigation framework with privacy filter. A point cloud map is processed to generate top-view and topological representations. Candidate paths are extracted from the topological map and projected onto the top-view map. A privacy filter perturbs human-related attributes, and the filtered information, together with the candidate paths, is provided to a vision–language model, which selects an optimal route under privacy constraints for robot navigation.

Theorems & Definitions (1)

• Remark 1.1