Real-time Optimal Navigation Planning Using Learned Motion Costs

Abstract

Navigation on challenging terrain topographies requires the understanding of robots' locomotion capabilities to produce optimal solutions. We present an integrated framework for real-time autonomous navigation of mobile robots based on elevation maps. The framework performs rapid global path planning and optimization that is aware of the locomotion capabilities of the robot. A GPU-aided, sampling-based path planner combined with a gradient-based path optimizer provides optimal paths by using a neural network-based locomotion cost predictor which is trained in simulation. We show that our approach is capable of planning and optimizing paths three orders of magnitude faster than RRT* on GPU-enabled hardware, enabling real-time deployment on mobile platforms. We successfully evaluate the framework on the ANYmal C quadrupedal robot in both simulations and real-world environments for path planning tasks on multiple complex terrains.

Learning Forward Dynamics Model and Informed Trajectory Sampler for Safe Quadruped Navigation

“…There were recent advances in the global planner using learning-based methods to find the global path faster [2,41,42,24,53] or to find a controller-aware path [14,52]. However, these methods still require a robust local planner to safely track the planned path.…”

Elevation Mapping for Locomotion and Navigation using GPU

“…This is useful in case of obstacle occlusion, as holes in the map need to somehow be interpreted by the downstream modules. A navigation planner, for example, would typically either optimistically fill the holes with an image inpainting algorithm [47] or simply consider them untraversable. The upper bound layer helps distinguish between safe patches caused by obstacle occlusions (which typically exhibit small inclinations in the upper bound) and unsafe larger drops which cause steeper ray angles resulting in larger inclinations.…”

“…We demonstrate our framework through extensive experiments. Our implementation supported legged locomotion and navigation research and formed the basis for perceiving surrounding terrains used by learning-based controllers [30], [38], [29], model-based controllers [23], [22], [16], or others such as navigation [45], [47] or learning occlusion filling [40]. In addition, the proposed elevation mapping solution was successfully used for underground exploration missions during Defense Advanced Research Projects Agency (DARPA) Subterranean Challenge [9] where team CERBERUS [8], [42] deployed our mapping on four quadrupedal robots, ANYmal [21].…”

Coupling Vision and Proprioception for Navigation of Legged Robots

“…Foothold scores can be estimated heuristically [14,18,32,40,52,80] or learnt [34,44,50,51,79]. Other methods forgo explicit foothold optimization and learn whether a given section of terrain is traversible [11,24,86]. Instead of using just vision, we combine planning module and locomotion policy via vision and proprioception.…”

“…sensor which is only capable of detecting clearly visible obstacles and regions that are hard to traverse, e.g. steps and ramps [14,80,83,86]. However, it is extremely challenging to predict several other terrain properties from vision like how slippery, uneven, granular or deformable is the surface.…”