PaperlayerView exact revision on arXiv

Source-linked AI summary

Estimating the Causal Impact of Social Navigation Signals in Corridor Scenarios

Francesco Baldo, Luca Castri, Gloria Beraldo, Nicola Bellotto

arXiv:2610.16411v1cs.ROcs.AI

TL;DR

Robots need to decide when dedicated social signals are relevant, but confounding can distort observation-based estimates of their effect. The paper models a corridor interaction causally and compares adjusted and naïve estimates. The naïve estimate suggests signalling harms success, whereas adjustment shows a positive effect.

  • Problem

    Robots need to determine when dedicated signals are relevant, but confounding can make observed relationships between signals and outcomes misleading.

    Source: §1 · pp. 1–2

  • Method

    The paper models a corridor interaction with a causal DAG and uses backdoor adjustment for the controlled static-obstacle confounder.

    Source: §3.2 · pp. 4–5

  • Results

    ATEnaïve = −0.21, while adjustment for static obstacles gives ATEcausal = +0.06, reversing the estimated effect of signalling.

    Source: §5.1 · pp. 7–8

  • Takeaways & Limitations

    Ignoring static-obstacle confounding can mislead a robot into suppressing a social signal that is beneficial after adjustment.

    Source: §6 · pp. 8–9

    The study has only 100 simulated episodes, sparse variable combinations, and deterministic pedestrian compliance.

    Source: §6 · p. 9

Abstract

from arXiv · show

Deploying robots in human-shared environments requires deciding not only how to navigate, but also when dedicated social signals are worth emitting. Classic observation-based inference can be misleading in such settings: confounding factors may bias the estimated relationship between a signal and the task outcome, leading to incorrect decisions. Causal inference provides a framework for disentangling these relationships by modelling the mechanisms linking observations, actions, and outcomes, and by estimating interventional effects after accounting for confounding paths. We study a human-robot interaction scenario in a corridor and use a causal model to estimate whether a social navigation signal causally improves transit success. We represent the interaction as a causal DAG involving robot action, pedestrian occupancy, available space, task outcome, and a controlled confounder (static obstacles). We compare the interventional estimate with the observational estimate obtained by conditioning only on the robot action, which does not explicitly model interventions. Our results show that accounting for confounding variables can reverse the conclusions drawn from naive observational analysis: the observational estimate wrongly suggests that the social signal is detrimental, whereas the causal estimate shows that it is beneficial. This discrepancy highlights how failing to account for confounding can lead to incorrect conclusions and demonstrates the importance of causal models for context-aware signalling decisions in social robot navigation.

1. Introduction

Human encounters rely on implicit motion cues, while robots often need dedicated signals whose relevance depends on context. This paper uses causal reasoning to test whether a signal improves interaction outcomes after accounting for confounding.

1. Introduction

  • Human motion conveys intent through changes in speed and direction, helping people resolve corridor encounters without words.A person is an active participant whose movement can communicate where they are going and what they are about to do.Source: §1 · p. 1
  • Robots cannot rely on human-like implicit communication, and dedicated intent signals are useful only when the robot can determine whether they are relevant.The paper frames causal reasoning as a way to distinguish genuinely relevant factors from misleading associations.Source: §1 · pp. 1–2
  • Confounding can make a person’s presence appear to require a signal even when the person is unaffected by the robot’s behaviour.Causal reasoning models links among observations, actions, and outcomes to support more context-dependent decisions.Source: §1 · p. 2
  • The paper asks whether emitting a social signal causally influences outcomes after contextual confounders are accounted for, rather than merely being associated with success.It models the interaction as a causal graph and compares causal and naïve analyses as a proof of concept.Source: §1 · p. 2

2. Related work

Prior robotics work applies causal models to several problems, but their use is uneven across domains. This paper instead uses causal analysis to assess the value of a specific social-navigation action.

2. Related work

  • Causal reasoning is adopted unevenly in robotics: causal concepts are often implicit in problem formulations and rarely stated or exploited.Robots can act differently in their environments, making interventional data available during ordinary operation.Source: §2 · p. 2
  • Prior work uses causal models for manipulation, planning under uncertainty, and failure handling, including state reduction, bias-robust policies, and corrective action selection.These approaches include interventions in an internal simulator, causal planning, and causal explanations of failures.Source: §2 · pp. 2–3
  • In human-robot interaction, prior work identifies states where robot behaviour influences nearby people, while this paper asks how much a specific action changes the interaction outcome.The earlier approach was evaluated with pedestrians simulated by the Social Force Model in a hallway.Source: §2 · p. 3
  • Other work applies causal models to human spatial behaviour and navigation or enables onboard causal analysis, whereas this paper compares naïve and adjusted estimates of one signal’s effect on task success.The comparison uses the same dataset for both estimates.Source: §2 · p. 3

3. Causal Impact Estimation of Social Navigation Signals

The study examines a robot traversing a pedestrian-shared corridor bottleneck and formalizes the interaction as a causal graph. The graph identifies static obstacles as a confounder that can be controlled through backdoor adjustment.

3.1. Human–Robot Spatial Interaction in a Corridor Scenario

  • At a T-shaped corridor junction, the robot chooses whether to request that pedestrians make way before crossing the bottleneck.Five pedestrian agents follow structured routines that create recurring spatial conflicts; the robot makes its decision at an observation waypoint.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2. Causal Impact Estimation

  • The simulation-specified DAG models the sequence from initial pedestrian occupancy through robot action and post-action occupancy to clearance and task outcome.Static obstacles occupy space but do not respond to the signal, unlike pedestrians; the model records six binary variables across the episode.Source: §3.2 · pp. 4–5
  • Static obstacles affect both the robot’s action choice and transit clearance, opening a backdoor path that conditioning on the recorded obstacle variable blocks.The set {O} satisfies the backdoor criterion, allowing the interventional effect to be identified by adjustment without blocking the directed action-to-outcome path.Source: §3.2 · p. 5

4. Experiments

The experiments use simulated corridor traversals to compare observational and backdoor-adjusted estimates of signal impact. The setup records episode variables and defines the two estimators, but these passages do not report their results.

4. Experiments

  • The study compares naïve and adjusted estimates of signalling’s effect on transit success using 100 simulated traversal episodes.Each episode is one traversal, and the two travel directions are analysed as a single dataset because their action and success rates are virtually identical.Source: §4 · p. 5, §4.1 · p. 7

4.1. Experimental Setup

  • The setup records one tuple of causal variables per episode, with deterministic pedestrian compliance and a response delay after signalling.The robot waits about five seconds before measuring post-action occupancy when it signals; the observational dataset contains 100 episodes.Source: §4.1 · pp. 6–7

4.2. Evaluation Metric & Comparison Baseline

  • The naïve baseline contrasts success probabilities conditioned on action, while the causal estimator computes interventional distributions by adjusting for static obstacles.The primary metric is the Average Treatment Effect on task success; the adjustment blocks the spurious path from action through obstacles to outcome.Source: §4.2 · p. 7

5. Results

The naïve observational estimate makes the social signal look harmful, but adjusting for static obstacles reverses the estimated effect. An alternative graph specification also yields a positive estimate, though the policy leaves one action-context combination unobserved.

5.1. Naïve vs. Causal Interventional Estimation

  • The comparison uses 100 recorded episodes, with the naïve estimate conditioning task success on robot action alone.That estimate is P(T=1|A=1)=0.63 versus P(T=1|A=0)=0.84.Source: §5.1 · p. 7
  • The estimated effect reverses from −0.21 observationally to +0.06 after backdoor adjustment for static obstacles.The naïve estimate conditions on robot action alone; the adjusted estimate weights contexts by the marginal distribution of obstacles.Source: §5.1 · pp. 7–8, Figure 6 · p. 8
  • Obstructed episodes were 16% of no-signal cases versus 50% of signal cases, while success with obstacles rose from 0.10 to 0.26 with signaling.The unequal context mix explains why pooling the episodes makes signaling appear associated with failure.Source: Figure 6 · p. 8, §5.1 · p. 8

5.2. Robustness of the Estimate

  • With an added Pi→Pe edge, adjustment on {Pi, O} still gives +0.11, and constrained hill-climbing recovers the added edge.The data cannot estimate the effect for clear-corridor cases in which the signal fires, because the policy never produces them.Source: §5.2 · p. 8
  • The policy never emits the signal in a clear corridor, leaving that action-context combination unobserved.The authors note that a small exploration probability could populate this cell.Source: §5.2 · p. 8

6. Conclusion

The paper concludes that adjusting for static-obstacle confounding reverses the naïve estimate of signaling’s effect on transit success. It also identifies limited sample coverage and a simulation-only setting as boundaries on the evidence.

6. Conclusion

  • Naïve conditioning suggests signaling lowers transit success by 21 percentage points, but adjustment for static-obstacle confounding reverses the estimated effect to a positive gain.Source: §6 · p. 8
  • Unadjusted decisions could lead a robot to suppress actions that are genuinely beneficial.Source: §6 · p. 9
  • The evidence is limited to 100 simulated episodes with deterministic pedestrian compliance, and some variable combinations are sparse or unobserved.The clear-corridor, signal-emitted case is empty because the policy never signals without perceived congestion.Source: §6 · p. 9

Declaration on Generative AI

The authors report using Claude for language editing and state that they reviewed the resulting text and retain responsibility for the publication.

Declaration on Generative AI

Related papers

Loading 2610.16411v1…