Interleaving symbolic and geometric reasoning for a robotic assistant

Samir Alili, Amit Kumar Pandey, Emrah Akin Sisbot, Rachid Alami2010

TL;DR

This paper presents a novel framework for robotic task planning that tightly integrates symbolic reasoning (what to do) with geometric reasoning (how to do it). By enabling a sophisticated interplay between these two planning domains, the system can generate more robust and feasible plans for complex tasks like human-robot cooperation, even when initial plans fail.

What Problem Does This Paper Address?

Traditional symbolic task planners, while powerful, often struggle when applied to real-world robotics. This is because they operate on abstract representations and can generate plans that are impossible to execute geometrically. For instance, a symbolic plan might say "pick up the cup," but it doesn't specify how to reach it without collision or where to place it for the human. This "symbolic-geometric gap" leads to plans that fail in practice. The authors aim to bridge this gap by creating a system where symbolic and geometric planners work hand-in-hand, allowing for continuous validation and refinement of plans at both levels. This is particularly crucial for cooperative tasks involving humans, where understanding reachability, visibility, and human capabilities is paramount.

Methods

The core of the proposed system is the "interleaving" of a symbolic planner (HATP - Human-Aware Task Planner) and a geometric planner.

  • Symbolic Planner (HATP): This is a hierarchical task network (HTN) planner that can generate plans for robots operating alongside humans. It considers not only the robot's capabilities but also the state and preferences of other agents (humans). HATP's plans are composed of actions and methods, which can be decomposed into simpler tasks.
  • Geometric Planner: This system handles the 3D environment, including obstacles, agents (robots and humans), and manipulable objects. It has three layers:
  • Geometric Tools Layer: Provides basic geometric analysis like reachability, visibility, and spatial relations.
  • Geometric Reasoning Layer: Finds specific points or configurations for basic actions (e.g., "put object for human to take") and plans agent paths.
  • Validation and Refinement Layer: Acts as an interface to the symbolic planner, receiving action requests and constraints, and returning feasibility information.
  • Interleaving Mechanism: The key innovation is how these two planners communicate. The symbolic planner proposes a plan, and for each action, it queries the geometric planner for feasibility. If the geometric planner finds a plan infeasible, it attempts to backtrack within its own domain to find alternative geometric solutions. If geometric backtracking fails, it informs the symbolic planner, which can then backtrack at the symbolic level to explore alternative actions or methods. This creates a feedback loop for continuous plan refinement.

Datasets

Not discussed in the paper. The paper focuses on a general framework and illustrates its functionality with a scenario.

Key Insights and Ideas

  • Bridging the Symbolic-Geometric Gap: The most significant insight is that a tight, iterative coupling between symbolic and geometric planners is essential for robust robotic task execution. Simply having separate planners is insufficient.
  • Geometric Backtracking: The ability of the geometric planner to backtrack and find alternative solutions before informing the symbolic planner is a crucial feature. This prevents unnecessary symbolic backtracking and allows for more efficient plan refinement.
  • Contextual Constraints: The symbolic planner can provide geometric constraints (e.g., "object must be visible to human") that refine the geometric planner's search space, leading to more context-aware plans.
  • Human-Awareness: The integration with HATP means the system can inherently consider human presence and capabilities, making it suitable for cooperative tasks.

Key Results

The paper demonstrates the framework's effectiveness through a detailed example of a coffee-making scenario involving a robot and a human. The example highlights how the system handles situations where initial geometric placements fail, leading to backtracking at both the geometric and symbolic levels to find a valid plan. While specific quantitative results (e.g., success rates, planning times) are not presented, the qualitative demonstration shows the system's ability to:

  • Generate a sequence of actions for a cooperative task.
  • Validate these actions geometrically, considering reachability and visibility for the human.
  • Backtrack at the geometric level to find alternative solutions when initial placements are not feasible.
  • Backtrack at the symbolic level if geometric solutions are exhausted.
  • Ultimately produce a valid plan that satisfies both symbolic and geometric constraints.

Limitations and Future Work

  • Limitations:
    • The paper focuses on a specific scenario and doesn't provide extensive quantitative evaluation across a range of tasks or environments.
  • The complexity of the backtracking mechanisms and their computational cost are not deeply analyzed.
  • The "mightability maps" mentioned in the context of human-robot interaction are introduced but not fully detailed in terms of their implementation and impact on planning.
  • Future Work:
    • Extending the framework to handle more complex human behaviors and preferences.
  • Developing more sophisticated cost-based plan enhancement strategies.
  • Further research into the interaction between symbolic and geometric backtracking to optimize efficiency.
  • Applying the framework to a wider variety of real-world robotic tasks.

Summary updated

Abstract

It is now well known that while symbolic task planners have been drastically improved to solve more and more complex symbolic problems the difficulty of successfully applying such planners to robotics problems still remains. Indeed, in
BibTeX
@misc{alili2010interleaving,
  author = {Samir Alili and Amit Kumar Pandey and Emrah Akin Sisbot and Rachid Alami},
  title = {{Interleaving symbolic and geometric reasoning for a robotic assistant}},
  year = {2010},
}

Loading related content...

Figure