Watch out! There may be a Human. Addressing Invisible Humans in Social Navigation

Phani Teja Singamaneni, Anthony Favier, Rachid Alami

TL;DR

Most social robots only react to people they can see, leading to "surprises" when someone rounds a corner. This paper introduces a method to detect "invisible humans"—potential people hidden behind occlusions—and integrates them as proactive constraints in a navigation planner, allowing robots to slow down and give corners a wide berth.

What Problem Does This Paper Address?

In social navigation, robots need to move in ways that are predictable and comfortable for humans. Current state-of-the-art planners are excellent at pathing around visible pedestrians. However, they suffer from a "blind spot" regarding occlusions.

When a human suddenly emerges from a doorway or around a sharp corner, the robot often reacts erratically, performing emergency stops or sharp swerves. This creates "shocks" or "surprises" for the human and makes the robot appear socially unintelligent. The authors tackle the challenge of making robots anticipate these hidden threats by treating occluded areas as locations where a human might be.

Methods

The authors propose a two-step approach: detection and integration.

  1. Invisible Human Detection: Using an emulated 2D laser scan on a known map, the system identifies "corners" (gaps in the laser contour larger than 0.5m). Using vector algebra and ray tracing, the algorithm calculates a point $H$ just out of sight where a human could realistically be standing.
  2. The "Invisible Humans Constraint": These predicted locations are fed into the CoHAN (Collaborative Human-Aware Navigation) framework. A new cost function is added to the robot’s optimization scheme. This cost considers:
  3. Average human walking speed (1.3 m/s).
  4. Human reaction time (0.5s).
  5. The robot's current velocity and distance to the occlusion.
  6. Narrow Passage Detection: The system uses the geometry of these "invisible humans" to identify doorways. If the robot and two predicted humans form an isosceles triangle within certain distance limits, the robot identifies a "narrow passage" and switches to a specific "Passing Through" modality, slowing down to $\leq 0.3$ m/s.

Datasets

The researchers evaluated their approach using:

  • Simulated Environments: They used the MORSE simulator with the InHuS (Intelligent Human Simulator) for realistic human behavior.
  • Maps: Testing was conducted on maps of the LAAS-CNRS lab, the Bremen environment, and a Random Maze generator to test robustness.
  • Real-World Hardware: Experiments were validated using a PR2 robot in a laboratory setting.

Key Insights and Ideas

  • Occlusions as Agents: The core "aha" moment is treating a geometric occlusion not just as a boundary, but as a "virtual agent" with a predicted velocity and reaction time.
  • Geometric Heuristics: Using a simple isosceles triangle to identify doorways is a computationally efficient way to trigger complex behavior changes (modality switching) without needing heavy machine learning models.
  • Balancing Act: The planner doesn't just avoid corners; it finds a "fine balance" between the forces exerted by visible humans and the potential forces of invisible ones. This prevents the robot from being "too conservative" and freezing in place.

Key Results

  • Detection Accuracy: The algorithm achieved a 76.85% accuracy in detecting valid emergence points across 200 experiments. If "overlap" detections (where the predicted human is slightly inside a wall) are included, accuracy rises to 89.16%.
  • Increased Safety Margin: In "sudden emergence" tests, the minimum distance maintained between the robot and a human increased significantly:
  • Standard Move Base: 0.584 m
    • CoHAN (visible only): 0.922 m
    • CoHAN with Invisible Human Constraint: 1.247 m
  • Qualitative Behavior: In real-world tests with a PR2, the robot successfully demonstrated "cautious" behavior, taking wider turns around corners and slowing down before entering doorways, matching the simulated results.

Limitations and Future Work

  • 2D Map Constraints: The system relies on a 2D laser scan. This can lead to false detections if the robot's sensors can see over an obstacle (like a table) that the base-level laser perceives as a solid wall.
  • False Positives in Complex Maps: In very cluttered environments, the algorithm can generate too many "invisible humans," potentially making the robot overly timid.
  • Future Directions: The authors plan to incorporate 3D sensor data to filter out false positives and refine the "Passing Through" modality to handle more complex social interactions in narrow corridors.

Summary updated

Abstract

Current approaches in human-aware or social robot navigation address the humans that are visible to the robot. However, it is also important to address the possible emergences of humans to avoid shocks or surprises to humans and erratic behavior of the robot planner. In this paper, we propose a novel approach to detect and address these human emergences called ‘invisible humans’. We determine the places from which a human, currently not visible to the robot, can appear suddenly and then adapt the path and speed of the robot with the anticipation of potential collisions. This is done while still considering and adapting humans present in the robot's field of view. We also show how this detection can be exploited to identify and address the doorways or narrow passages. Finally, the effectiveness of the proposed methodology is shown through several simulated and real-world experiments.
BibTeX
@inproceedings{singamaneni2022watch,
  author = {Phani Teja Singamaneni and Anthony Favier and Rachid Alami},
  title = {{Watch out! There may be a Human. Addressing Invisible Humans in Social Navigation}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  doi = {10.1109/iros47612.2022.9982186},
}

Loading related content...

Figure