Heralding the advent of autonomous vehicles and mobile robots that interact with humans, responsibility in spatial interaction is burgeoning as a research topic. Even though metrics of responsibility tailored to spatial interactions have been proposed, they are mostly focused on the responsibility of individual agents. Metrics of causal responsibility focusing on individuals fail in cases of causal overdeterminism - when many actors simultaneously cause an outcome. To fill the gaps in causal responsibility left by individual-focused metrics, we formulate a metric for the causal responsibility of groups. To identify assertive agents that are causally responsible for the trajectory of an affected agent, we further formalise the types of assertive influences and propose a tiering algorithm for systematically identifying assertive agents. Finally, we use scenario-based simulations to illustrate the benefits of considering groups and how the emergence of group effects vary with interaction dynamics and the proximity of agents.
Figures & tables
Figure 1: Feasible action space reduction [ 17 ] : For the robot crossing (a) represented in the grid world (b), the feasible action space reduction (FeAR) imposed by actors on affected agents are computed based on the feasible actions of the affected when actors follow their Move de Riguer (MdR) (represented by ) and how many of these are rendered infeasible by the actual actions of actors (represented by ). For affected agent 2, (c) shows how agent 1 reduces the feasible action space by two, (d) shows how 5 on its own has no influence, and (e) shows how the group {5,7} reduces the feasible action space by two.
Figure 2: Types of assertive influence based on group FeAR: For the illustrative scenario (a) where three agents are moving towards each other, we show how counterfactuals based on the MdR of actors (b) are used to compute iFeAR values (c) for individual actors and gFeAR values for group actors (d). While iFeAR can only identify solo influences ( 1\nolinebreak↼\nolinebreak2 and 2\nolinebreak↼\nolinebreak1 ) (e), analysing gFeAR can reveal additional mediated ( 2\nolinebreak↼1\nolinebreak3 ) (f) and coupled ( 3\nolinebreak⇐\nolinebreak{1,2} ) (g) influences. Even though FeAR3,3<1 reflects the reduction in feasible action space of 3, iFeAR cannot identify the assertive actors; which are revealed by gFeAR ( 3\nolinebreak⇐\nolinebreak{1,2} ). Also note that agent 3 has no influence on agent 1 as FeAR2,1=FeAR{2,3},1 .
\begin{array}[]{l r}\phi_{j}\leftarrow\{i:\mathrm{FeAR}_{i,j}<0\}&\text{Find courteous agents \phi_{j}.}\\ \kappa_{j}\leftarrow\neg j\setminus\phi_{j}&\text{Remove \phi_{j}fromcandidateagents\kappa_{j}.}\\ \mathcal{R}_{0}\leftarrow\emptyset&\text{\mathcal{R}{n}−−−Assertiveagentsuntil\mathbb{T}{j,n}.}\\ \textbf{for }n=1,2,\dots:&\text{For each tier,}\\ \hskip 8.19447pt\mathbb{T}_{j,n}\leftarrow\emptyset&\\ \hskip 8.19447pt\textbf{for }k=1,\dots,|\kappa_{j}|:&\text{starting with k=1,}\\ \hskip 8.19447pt\hskip 8.19447pt\textbf{for all }G\subseteq\kappa_{j},|G|=k:&\text{loop through groups with k agents,}\\ \hskip 8.19447pt\hskip 8.19447pt\hskip 8.19447pt\textbf{if }\mathrm{FeAR}_{\mathcal{R}_{n-1}\cup G,j}>\mathrm{FeAR}_{\mathcal{R}_{n-1},j}:&\text{identify assertive groups,}\\ \hskip 8.19447pt\hskip 8.19447pt\hskip 8.19447pt\hskip 8.19447pt\mathbb{T}_{j,n}\leftarrow\mathbb{T}_{j,n}\cup\{G\},&\text{add them to tiers, and}\\ \hskip 8.19447pt\hskip 8.19447pt\hskip 8.19447pt\hskip 8.19447pt\kappa_{j}\leftarrow\kappa_{j}\setminus G&\text{remove them from candidate agents.}\\ \hskip 8.19447pt\textbf{if }\mathbb{T}_{j,n}=\emptyset:\textbf{ break}&\\ \hskip 8.19447pt\mathcal{R}_{n}\leftarrow\mathcal{R}_{n-1}\cup\bigcup_{G\in\mathbb{T}_{j,n}}G&\end{array}
Algorithm 1 Tiering the assertive influence of agents
Figure 4: Ranking the assertiveness of agents into tiers Tj,n : In this illustrative scenario (a), when considering agent 1 as the affected, iFeAR only show agent 6 as being the assertive (b). However, counterfactuals with groups (c) reveal more assertive influences on agent 1 which are systematically ranked into tiers T1,n (d).
Figure 6: S1: Uncovering group effects with group FeAR: For the robot crossing scenario from Fig. 1 , represented in the grid world as in (a), compared to just using iFeAR (b), we are able to uncover more assertive influences using gFeAR, either through the Shapley values (c) or tiers (d). For example, while iFeAR only shows the assertive influence of the robot 5 on pedestrians 3, 4 and 6, both Shapley values and tiers show that 5 is assertive towards all the pedestrians.
Table 1: Ranks of assertive influence in S1: Ranking of assertive influence calculated using FeAR, tiers from group FeAR and Shapley values of group FeAR. ‘ ■ ’ show affected agents which are not considered in the rankings.
Figure 8: Emergence of group effects in randomised simulations: Different simulation scenarios are shown in (a). Difference in the number of assertive agents ΔjAssertive identified using individual FeAR (iFeAR) and (tiers of) group FeAR (gFeAR) for different proximity to the affected agent are shown in (b). Rankings made using FeAR, tiers from gFeAR or Shapley values of gFeAR we compared using Kendall’s τ . Variation of Kendall’s τ with respect to the scenarios and median Manhattan distance between agents is shown in (c).
Human behavior in interactive settings is shaped not only by individual objectives but also by shared constraints with others, such as safety. Understanding how people allocate responsibility, i.e., how much one deviates from their desired policy to accommodate others, can inform the design of socially compliant and trustworthy autonomous systems. In this work, we introduce a method for learning a probabilistic responsibility allocation model that captures the multimodal uncertainty inherent in multi-agent interactions. Specifically, our approach leverages the latent space of a conditional variational autoencoder, combined with techniques from multi-agent trajectory forecasting, to learn a distribution over responsibility allocations conditioned on scene and agent context. Although ground-truth responsibility labels are unavailable, the model remains tractable by incorporating a differentiable optimization layer that maps responsibility allocations to induced controls, which are available. We evaluate our method on the INTERACTION driving dataset and demonstrate that it not only achieves strong predictive performance but also provides interpretable insights, through the lens of responsibility, into patterns of multi-agent interaction.
Isaac Remy, Caleb Chang, Karen Leung
University of Washington, Department of Aeronautics and Astronautics · NVIDIA
With the growing adoption of artificial intelligence in high-stakes decision-making, identifying the causes of outcomes--particularly failures--and determining who is responsible has become a critical concern. In this work, we examine how well formal definitions of responsibility attribution, grounded in the framework of actual causality, align with human judgments of responsibility. To this end, we conduct a large-scale survey to elicit human judgments of responsibility in multi-agent sequential decision-making scenarios, using a modified version of the card game Goofspiel. We evaluate multiple responsibility attribution methods, assess their alignment with human judgments about responsibility, and identify factors that significantly shape responsibility judgments. While no single responsibility attribution method consistently aligns with human responses, our findings highlight key factors that influence human responsibility judgments, including agent-specific biases and amount of information available to agents during decision-making.
Nripsuta Ani Saxena, Stelios Triantafyllou, Goran Radanović
University of Southern California Los Angeles, California, USA · Max Planck Institute for Software Systems Saarbrücken, Germany
Physical automation is scaling toward fleets of embodied machines commanded by an AI brain. Early deployments already run factories and warehouses at production rates beyond any human line, and their adoption is accelerating. But when their joint decisions cause harm, everyone involved has reason to blame everyone else, the machine vendor, the algorithm provider, the factory operator, the insurer, and the regulator, and no method can divide the responsibility between them. Existing methods read logs whose origin they cannot verify and name a single culprit, misrepresenting outcomes that are overdetermined, preempted, or caused by an omission. We present AUDITA, an audit layer pairing a tamper-evident record of every inter-agent command with a certified, graded causal-attribution engine. We prove its verdict cannot be gamed: a rule-following agent can never be made to look guilty, an attempt to shift blame is itself caught and graded, and we establish the exact limit of what an evidence-based auditor can certify. On live language-model pipelines it reduces the standard judge baseline's responsibility error roughly threefold; on a benchmark of accident-grounded structures it recovers responsibility where single-culprit baselines fail, and stays invariant under forgery. AUDITA turns the question of who is to blame from an argument about logs into a calculation over evidence.
Zhixu Du, Yiran Chen
Department of Electrical and Computer Engineering, Duke University, Durham, North Carolina, USA