Learning Spatial Constraints using Gaussian Process for Shared Control of Semi-autonomous Mobile Robots

Kun Qian, Dan Niu, Fang Fang, Xudong MaICINCO 2017 · DOI

TL;DR

This paper introduces a novel shared control system for semi-autonomous mobile robots that learns and enforces spatial constraints, specifically maintaining a reliable Wi-Fi connection. By using Gaussian Processes to map Wi-Fi signal strength (WSS) and fusing it with occupancy grids, the robot can autonomously avoid low-signal areas, enhancing teleoperation reliability.

What Problem Does This Paper Address?

Tele-operated mobile robots are increasingly used for tasks like environment inspection and monitoring. However, in indoor environments, maintaining a stable Wi-Fi connection is crucial for reliable control and high-quality feedback. When a robot enters areas with weak Wi-Fi signals, teleoperation can become unreliable, leading to loss of control or degraded sensory input. This paper addresses the challenge of ensuring continuous and reliable Wi-Fi connectivity for a tele-operated mobile robot by developing a "WSS-aware" navigation behavior. This is framed as a spatially restricted navigation problem, where the robot must not only avoid physical obstacles but also adhere to the constraint of maintaining adequate Wi-Fi signal strength.

Methods

The core of the proposed solution lies in a two-pronged approach:

  1. WSS Distribution Learning: The paper employs Gaussian Process Regression (GPR) to learn the spatial distribution of Wi-Fi Signal Strength (WSS) within an indoor environment. GPR is chosen for its ability to model complex, non-linear relationships and provide uncertainty estimates, which is crucial for predicting WSS at unmeasured locations. The robot collects WSS samples as it explores and builds an occupancy grid map simultaneously using GMapping. These WSS samples, along with their corresponding locations, form the training data for the GPR model.
  2. WSS-Occupancy Hybrid Mapping and Shared Control: The learned WSS distribution is then fused with the robot's occupancy grid map to create a WSS-Occupancy hybrid map. This hybrid map visually represents both physical obstacles and areas with low Wi-Fi signal strength. A shared control strategy is implemented to leverage this hybrid map. In this strategy, the robot's autonomous navigation system generates a path that avoids low-WSS areas. This autonomous command is then combined with human teleoperation commands using a weighting mechanism. The system dynamically adjusts the balance between human control and autonomous avoidance based on the predicted safety of the human command relative to the WSS constraint.

Datasets

The paper does not explicitly mention using pre-existing public datasets. Instead, it describes experiments conducted in two settings:

  • Simulation: A Stage-based robot simulator was used to generate training data for the WSS-Occupancy hybrid map.
  • Real-world Experiments: Experiments were conducted using a Turtlebot mobile robot in an office environment.

The data collected in these experiments (robot trajectories, Wi-Fi signal strength measurements, and occupancy grid maps) served as the basis for training and evaluating the proposed system.

Key Insights and Ideas

  • Wi-Fi as a Spatial Constraint: The paper effectively reframes Wi-Fi signal strength not just as a communication quality metric but as a critical spatial constraint that must be factored into robot navigation.
  • Gaussian Processes for Environmental Modeling: GPR proves to be a powerful tool for learning and predicting the spatial distribution of non-uniform environmental features like Wi-Fi signals, especially with limited training data.
  • Hybrid Mapping for Enhanced Navigation: Fusing Wi-Fi signal strength information with traditional occupancy grids creates a richer environmental representation that enables more intelligent path planning.
  • Shared Control for Robust Teleoperation: The proposed shared control framework effectively balances human intent with autonomous safety measures, ensuring that teleoperation remains reliable even when facing environmental constraints.

Key Results

  • Successful WSS Mapping: The GPR model successfully learned a smooth 2D spatial distribution of WSS from limited training samples, accurately predicting signal strength in unmeasured areas.
  • Effective WSS-Aware Path Planning: The WSS-Occupancy hybrid map enabled the robot to generate paths that actively detoured around areas with poor Wi-Fi signal strength, as demonstrated in simulation and real-world experiments.
  • Guaranteed Wi-Fi Connectivity: Experiments showed that WSS-aware navigation behavior guaranteed continuous and reliable Wi-Fi connection, validated by comparing WSS readings along different paths.
  • Robust Shared Control: Real-world experiments demonstrated that the shared control system could adapt its behavior, prioritizing autonomous avoidance of low-WSS areas when human commands risked violating the constraint. In one scenario, it successfully navigated a narrow corridor by switching to full human supervision when autonomous planning failed.

Limitations and Future Work

  • Static Constraint Assumption: The current approach treats the Wi-Fi distribution as static. Future work could explore dynamic WSS environments where Wi-Fi access points might change their locations or signal strengths.
  • Generalization of Constraints: While this paper focuses on Wi-Fi, the underlying framework of learning spatial constraints and using shared control could be extended to other types of environmental constraints (e.g., temperature, humidity, safety zones).
  • Human Supervision Effectiveness: Further investigation into the optimal balance between human control and autonomous intervention in various scenarios could enhance the system's adaptability.
  • Real-time Performance: While not explicitly detailed, optimizing the GPR inference and fusion process for real-time navigation in complex, dynamic environments would be a valuable future direction.

Summary updated

BibTeX
@inproceedings{qian2017learning,
  author = {Kun Qian and Dan Niu and Fang Fang and Xudong Ma},
  title = {{Learning Spatial Constraints using Gaussian Process for Shared Control of Semi-autonomous Mobile Robots}},
  booktitle = {ICINCO},
  year = {2017},
  doi = {10.5220/0006382902010208},
}
Figure