Dynamo

Fast Online Learning for Adaptive Mobility

Publications

• Brian Yamauchi, “Fast Online Learning of Control Regime Transitions for Adaptive Robotic Mobility,” Proceedings of SPIE Vol. 8387: Unmanned Vehicle Technology XIV, April 2012.
  Download

Video

• Dynamo Overview
• Learning Rollover Prevention for the PackBot
• Learning Active Weight Shifting for the PackBot
• Learning Understeer/Oversteer Prevention for Simulated Race Cars in TORCS

The Need for Speed

• Faster UGVs will be useful in a wider range of infantry missions
• Faster UGVs will minimize the time warfighters are exposed to danger

The Need for Rough Terrain Mobility

• Warfighters need to fight in rugged terrain, and UGVs will need to reliably traverse this terrain

The scientific goal of the Dynamo Project is to:

• Develop learning techniques that will allow UGVs to adapt quickly to a wide range of non-linear vehicle control regimes

The application goal of the Dynamo Project is to:

• Develop driver-assist behaviors that will aid operators in control of small UGVs in highly-dynamic environments including:
• High-speed control of wheeled UGVs
• Control of tracked UGVs over rough terrain

We have developed a new, online, real-time learning algorithm that we call Dynamic Threshold Learning (DTL) that allows robots to quickly learn the non-linear transitions between different control regimes.  For example, DTL can learn the steering and acceleration inputs that may cause a wheeled vehicle to lose control during aggressive maneuvers.  A DTL-enabled controller can then intervene to prevent the driver from losing control.  Likewise, DTL can learn the pitch and roll conditions that indicate a tracked vehicle is about to roll over in rough terrain.  A DTL-enabled controller can then modulate the vehicle commands to prevent the vehicle from rolling over.