Learning Robust Plans for Mobile Robots from a Single Trial

AAAI 1996

TL;DR

This paper introduces a method for learning robust, reactive robot navigation plans from a single observed behavior trace. By translating sequences of "behavior events" into a reactive plan language, the system can generalize from a single example, making plans resilient to sensor noise and minor environmental changes. This approach allows robots to learn from incidental experiences, not just explicitly planned ones.

What Problem Does This Paper Address?

Developing intelligent mobile robots requires them to navigate complex environments. Traditional planning methods often rely on hand-crafted plans or extensive training data. While sophisticated "reactive" planning formalisms exist, creating plans that are robust to real-world uncertainties like sensor noise or slight environmental variations is challenging. Existing case-based learning approaches typically require a library of pre-existing plan fragments, limiting learning to situations similar to those already encountered and precluding serendipitous learning from unexpected behaviors. This paper tackles the problem of learning effective, robust plans from a single observed instance of desired behavior, enabling robots to learn from their experiences more efficiently and adaptively.

Methods

The core of the paper's method is trace translation. The system observes a robot's behavior, which is represented as a sequence of "behavior events." These events are qualitative changes in the state of the robot's underlying low-level control system (e.g., a wall-following process signaling completion). The system then translates this sequence of events into a reactive plan written in a subset of the RPL (Reactive Plan Language).

Key aspects of the method include:

  • Event-based Representation: Behavior is broken down into discrete, qualitative events, which are easier to reason about than continuous sensor streams.
  • Translation Rules: A set of rules maps specific event specifications (process, event type, parameters) to RPL plan fragments.
  • Variabilization: The system identifies and uses variables to link events, allowing plan steps to depend on values produced earlier in the trace. It assumes equal values are logically identical.
  • Post Facto Parameterization: To handle noise, the system searches the trace for needed parameters to make plan steps more robust. For example, if a command to turn 90 degrees actually resulted in a 94.6-degree turn, the learned plan will incorporate this observed value.
  • Preprocessing: The system includes steps to handle "dangling activations" (behaviors that start but aren't explicitly signaled as starting in the trace) and "irrelevant events" (actions that don't affect the outcome and could introduce unreliability).

Datasets

The paper evaluates its system using the ARS MAGNA mobile robot simulator. This simulator provides a 2D environment with walls and various objects, and the robot is equipped with basic behaviors (like "follow wall," "go to designator") and sensing processes (like "acquire designator"). The simulator incorporates noise into both behavior processes and sensing processes, making it suitable for testing the robustness of learned plans. The paper does not introduce new benchmarks but uses this existing simulation environment.

Key Insights and Ideas

  • Learning from "Incidental" Experience: The most significant insight is the ability to learn useful plans from behaviors that might not have been explicitly intended or even recorded as part of a planned task. This opens up a new avenue for robot learning beyond supervised or goal-directed training.
  • Robustness through Event Abstraction and Post-Hoc Parameterization: By focusing on qualitative "behavior events" and incorporating observed deviations (post-hoc parameterization), the learned plans become inherently more resilient to noise and minor environmental changes than plans derived solely from ideal commands.
  • Generalization from a Single Trial: The core contribution is demonstrating that a single, observed behavior trace can be sufficient to learn a plan that reliably repeats that behavior under similar conditions, a significant step towards more efficient robot learning.

Key Results

The paper demonstrates the effectiveness of its trace translation method on mobile robot navigation tasks within the ARS MAGNA simulator.

  • Hand-Generated Behavior: A manually generated behavior trace resulted in a learned plan that successfully repeated the robot's trajectory in 9 out of 10 test runs, even with added obstacles. The few failures were attributed to perceptual noise that would be difficult to avoid without more domain knowledge.
  • Automatically Generated Behaviors: When using traces from a random wandering plan, the learned plans also showed good performance. For one experiment, 9 out of 10 runs succeeded despite variations in the robot's initial trajectory. In another, the learned plan achieved the goal in 7 out of 10 runs, even with obstacles in different positions, showcasing robustness to environmental changes.
  • Qualitative Findings: The learned plans reliably repeated the original behavior, exhibiting variance in specific trajectories due to noise but achieving the same overall effect. The system's ability to handle obstacles and generalize across different environmental configurations was highlighted.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Timing: The current plans are purely reactive and lack explicit notions of time. This can lead to issues if a significant time gap between events in the trace is crucial for the robot's position or the world state. Future work could involve adding a "timing behavior process" or using odometry to measure movement between events.
  • Effector Failure: While the system aims for robustness, effector failures (like losing a designator) can still cause plans to fail. The authors suggest using "policies" (a construct in RPL) to maintain needed conditions and recover from failures, potentially by designing heuristic policies for common plan steps.
  • Global Understanding: The current approach assumes no knowledge beyond the translation rules, meaning it lacks a deep understanding of the complete context of each event. Future work proposes using the original plan that generated the behavior trace as a resource to abstract up the plan hierarchy and create more robust, generalized plans.
  • Indexing and Evaluation: The paper notes that properly indexing new plans in a case base and evaluating their usefulness are beyond its scope but are important related problems.
  • Recognizing Interesting World States: Identifying when a world state is "interesting" enough to warrant learning is also an open problem.

Summary updated

BibTeX
@inproceedings{1996learning,
  title = {{Learning Robust Plans for Mobile Robots from a Single Trial}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure