Experiments

• Combined Collision Detection and Avoidance

  We conducted a series of experiments with the Kodiak robot tracking a path and detecting and avoiding obstacles in the spring of 2007 at the Robot City test facility in Pittsburgh. The vehicle logged over 100 kilometers of autonomous driving, with the navigation system achieving a successful obstacle avoidance rate of over 95%. When the vehicle was not able to avoid an obstacle it stopped in front, rather than colliding. Here are a few movies from these tests:

  A complete loop with max speed 5 m/s
  
  Vehicle with a wide obstacle and a slalom scenario
  
  Visualization of the Dodger algorithm, data logged from the complete loop test.
  
  
  
• Machine Learning to Improve Collision Avoidance

  The initial implementation of the Dodger algorithm used a set of hand-tuned parameters to control the strength of attraction to the goal and repulsion from obstacles. Later we used machine learning to derive a new set of parameters. We tested the behavior of the system using the different parameter sets by running the Kodiak robot through a series of obstacle avoidance scenarios of varying complexity. The autonomous system using the learned parameter set showed a significant improvement, with an overall successful collision avoidance rate of 92% using the learned set, compared with a 65% success rate using the hand-tuned set.

  

  The images above show sample paths driven by the autonomous system using the learned parameters (purple) and the hand-tuned parameters (green). Using the learned parameters, the vehicle turns sooner and gives the obstacles wider berth.