A Natural Interface and Unified Skills for a Mobile Robot

AAAI 1999

TL;DR

This paper presents a novel approach to mobile robot interaction and navigation by integrating a natural language and gesture interface with a unified set of navigation skills. The system allows for more intuitive human-robot communication and robust navigation in unknown environments by combining exploration, localization, and planning using evidence grids.

What Problem Does This Paper Address?

Interacting with mobile robots in complex, dynamic, and unknown environments poses significant challenges. Traditional command interfaces are often rigid and require users to be overly explicit about every action. Furthermore, robots need to navigate effectively in environments they haven't encountered before, which requires robust mapping, localization, and path planning capabilities that can adapt to changes. This paper tackles the problem of creating a more natural and intuitive way for humans to command robots, while simultaneously equipping the robot with the necessary skills to autonomously explore, map, and navigate these challenging, unfamiliar spaces.

Methods

The paper proposes a two-pronged approach:

  • Natural Interface: This component combines natural language and gestures for human-robot communication. It uses "context predicates" to track the state of various goals during interaction, allowing the robot to understand and pursue unaccomplished goals even if commands are interrupted or incomplete. This enables more flexible and less explicit user interaction.
  • Unified Navigation Skills: The core of the navigation system relies on evidence grids as a unifying representation. This allows for the integration of several key skills:
  • Exploration: A frontier-based strategy is used to explore unknown environments, expanding the robot's knowledge by moving towards boundaries between known and unknown areas.
  • Continuous Localization: This method combats odometry drift by continuously correcting the robot's position estimates using sensor data.
  • Mapping: Two types of maps are maintained: short-term perception maps (for immediate sensor data) and long-term metric maps (representing the environment over time).
  • Planning: A propagation-based path planner (Trulla) uses the long-term map to compute navigation paths.
  • Reactive Navigation: The Vector Field Histogram (VFH) method is integrated to handle transient obstacles not yet represented in the maps, using the short-term perception map for immediate obstacle avoidance.

The integration of these skills is facilitated by a "Map Server" that makes sensor-fused perceptions available to different components.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Synergy of Modalities: Combining natural language and gestures for robot commands is more natural and effective than using either alone, mirroring human communication patterns.
  • Context is King: Tracking "context predicates" is crucial for enabling robots to understand and complete goals autonomously, even with interrupted or fragmented commands. This allows for more independent robot operation.
  • Evidence Grids as a Unifying Framework: Using evidence grids as a common representation significantly simplifies the integration of diverse navigation skills like exploration, localization, planning, and reactive avoidance.
  • Adaptive Navigation: The system's ability to adapt its long-term map to environmental changes and combine global planning with reactive obstacle avoidance makes it robust in dynamic and unknown environments.

Key Results

The paper describes a functional system that has been tested with a mobile robot. While specific quantitative results (e.g., navigation success rates, localization accuracy percentages) are not detailed, the authors state that the system allows a robot to:

  • Enter an unknown environment and autonomously map it.
  • Maintain accurate odometry through continuous localization.
  • Robustly plan and navigate in real-time within the environment.
  • Adapt its maps to changes in the environment.
  • Avoid both static and transient obstacles.

The architecture is illustrated in Figure 1, showing the flow of information between the different components.

Limitations and Future Work

  • Limitations: The paper notes that the continuous localization's long-term map update method can adapt to somewhat rapid and persistent changes but not to very fast changes (e.g., a person walking through a room). This means that paths generated by Trulla might not be sufficient to prevent collisions with such transient obstacles.
  • Future Work: The authors express a desire to extend goal tracking so that the mobile robot can complete semantically related goals that are not initially specified or are unknown to the human at the time the initial goal is instantiated. This suggests a move towards more proactive and inferential goal completion by the robot.

Summary updated

BibTeX
@inproceedings{1999natural,
  title = {{A Natural Interface and Unified Skills for a Mobile Robot}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure