A Practical Approach to Learning Dynamics for Rough Terrain Navigation

  Unmanned ground vehicles offer great potential, but struggle in rough terrain environments  due to their inability to reason about complex dynamics over longer horizons. Thus, driving  strategies are often short-sighted and underutilize the robot’s dynamic capabilities. This  thesis addresses the challenges of long-horizon, dynamics-aware decision making in rough  terrain, arising from the inability to model complex system dynamics. The findings reveal  two key insights: 1) leveraging low-quality simulation data can reduce training requirements  of real-world dynamics models; and 2) long-horizon decision making can still utilize imprecise  dynamics models through careful handling of prediction uncertainty.  

We first discuss our model-based reinforcement learning approach for rough terrain navigation. This approach trains a dynamics model to predict the robot’s trajectory over uneven  terrain and capture prediction uncertainty. Decision making uses this model along with a  closed-loop divergence constraint to aid in longer-horizon trajectory prediction and prevent  exploitation of potentially problematic modeling errors. We show that this approach leads  to robust, non-myopic driving strategies that take full advantage of the robot’s capabilities.  

Next, we explore leveraging simulation to reduce real-world training data requirements.  This culminates in an approach that uses a large variety of simulated systems to train  a dynamics model that quickly and probabilistically adapts to any new, including real world, target system using any available target system data. Using this model within an  uncertainty-aware decision making framework results in safe, albeit low performance, driving  upon initialization. As more target system observations are collected, the adaptive dynamics  model becomes more tailored to the target system resulting in increased driving performance.  

Finally, we combine these concepts to form a non-myopic rough terrain navigation frame work that can quickly and robustly adapt to new target systems. We show that upon ini tialization, this framework chooses conservative routes that avoids obstacles. However, after  just one demonstration of driving over obstacles, the framework chooses more aggressive  routes over obstacles that match the system’s capabilities.