Learning obstacle avoidance parameters from operator behavior

Bradley Hamner, Sanjiv Singh, Sebastian SchererJournal of Field Robotics 2006 · DOI

TL;DR

This paper presents a method for an outdoor mobile robot to learn how to avoid obstacles by observing human drivers. By analyzing how humans navigate different obstacle densities and driving styles (careful vs. aggressive), the robot automatically tunes its obstacle avoidance parameters, leading to more human-like and effective navigation.

What Problem Does This Paper Address?

Autonomous outdoor robots, especially those operating at high speeds, need to navigate complex environments with unpredictable obstacles while still following a designated path. Traditional methods often rely on hand-tuned parameters for obstacle avoidance, which is a tedious and time-consuming process. These hand-tuned parameters may not generalize well to different driving styles or environmental conditions (e.g., sparse vs. dense obstacles). The challenge is to create a system that can adapt its obstacle avoidance behavior to mimic human drivers, who naturally exhibit varying levels of caution and aggressiveness based on the situation.

Methods

The core of the paper is the adaptation and learning of the Fajen and Warren (MFW) obstacle avoidance model. This model uses a reactive approach, combining attraction to a goal point with repulsion from obstacles to generate steering commands. The authors identify five key parameters in this model that control the balance between following the path and avoiding obstacles.

Their novel contribution lies in learning these parameters automatically. They achieve this by:

  • Human Driving Data Collection: A human subject drives a modified all-terrain vehicle (ATV) equipped with sensors. The robot's environment (goal point, obstacles) and the human's path are recorded.
  • Behavioral Categorization: The collected data is categorized into four distinct driving styles based on obstacle density (sparse/dense) and aggressiveness (careful/aggressive).
  • Parameter Optimization: For each category, the authors use a two-step optimization process. First, they generate random parameter sets and select the best ones. Then, they use a nonlinear least squares method to fine-tune these parameters by minimizing the difference (Euclidean distance) between the robot's simulated path (generated using the MFW model with the current parameters) and the human driver's recorded path.
  • Speed Control: A separate speed control function is introduced, which is also dependent on obstacle distance and angle, to ensure safe operation at high speeds.

Datasets

The paper describes data collection using a modified all-terrain vehicle (ATV) equipped with laser range finders and a high-accuracy GPS/IMU system. The training data consists of twelve paths driven by an operator through various obstacle configurations (dense and sparse) and driving styles (careful and aggressive). These paths were manually selected for segments exhibiting obstacle avoidance behavior. The authors do not explicitly mention a separate benchmark dataset for evaluation, but they test the learned parameters on new obstacle situations with the ATV.

Key Insights and Ideas

  • Human Driving as a Gold Standard: The paper strongly advocates for using human driving behavior as the target for autonomous robot navigation. Humans naturally adapt their driving style to the environment, and learning from them provides a robust way to achieve similar adaptability.
  • Parameterization is Key: The MFW model's effectiveness hinges on its parameters. Automatically learning these parameters is crucial for achieving good performance across different scenarios.
  • Context Matters: Different driving situations (dense vs. sparse obstacles, careful vs. aggressive intent) require distinct parameter sets. A single set of parameters is unlikely to perform optimally in all conditions.
  • Balancing Act: The core challenge is balancing path following (attraction to the goal) with obstacle avoidance (repulsion from obstacles). The learned parameters effectively tune this balance for specific contexts.
  • Speed and Safety Trade-off: High-speed operation necessitates careful consideration of obstacle avoidance. The learned parameters allow the robot to be appropriately cautious or aggressive depending on the learned style.

Key Results

  • The learned parameter sets for each of the four categories (Careful/Sparse, Careful/Dense, Aggressive/Sparse, Aggressive/Dense) generally performed better than a hand-tuned set in their respective categories.
  • The "Careful/Sparse" parameter set provided wide berth with low oscillation in sparse environments.
  • The "Careful/Dense" set, while safe, exhibited large oscillations in sparse environments due to high goal and obstacle weights.
  • The "Aggressive/Sparse" set showed poor obstacle avoidance, suggesting insufficient training data or an overly aggressive operator style.
  • The "Aggressive/Dense" set was aggressive but tracked a safe path, demonstrating good performance in dense environments.
  • The learned parameter sets, particularly AD and CD, navigated safely through both dense and sparse configurations, outperforming the hand-tuned set in dense scenarios.

Limitations and Future Work

  • Training Data Selection: The current method relies on manually selected path segments for training. The authors suggest that an automatic method for selecting relevant path segments would improve generalization.
  • Online Classification: The system requires an online method to classify instantaneous obstacle configurations as dense or sparse to select the appropriate parameter set. This is identified as future work.
  • Oscillation Reduction: The authors believe oscillation could be further reduced by introducing an additional metric during training that penalizes frequent steering angle changes.
  • Generalizability of the MFW Model: While the MFW model is effective, its applicability to other obstacle avoidance algorithms or different robot platforms is not explicitly discussed.
  • Aggressive/Sparse Performance: The poor performance of the aggressive/sparse parameter set highlights a potential need for more diverse training data or a more sophisticated learning approach for extreme driving styles.

Summary updated

Abstract

Abstract This paper concerns an outdoor mobile robot that learns to avoid collisions by observing a human driver operate a vehicle equipped with sensors that continuously produce a map of the local environment. We have implemented steering control that models human behavior in trying to avoid obstacles while trying to follow a desired path. Here we present the formulation for this control system and its independent parameters and then show how these parameters can be automatically estimated by observing a human driver. We also present results from operation on an autonomous robot as well as in simulation, and compare the results from our method to another commonly used learning method. We find that the proposed method generalizes well and is capable of learning from a small number of samples. © 2007 Wiley Periodicals, Inc.
BibTeX
@article{hamner2006learning,
  author = {Bradley Hamner and Sanjiv Singh and Sebastian Scherer},
  title = {{Learning obstacle avoidance parameters from operator behavior}},
  journal = {Journal of Field Robotics},
  year = {2006},
  doi = {10.1002/rob.20171},
}
Figure