Real-time Optimal Navigation Planning Using Learned Motion Costs

Navigation on challenging terrain requires the understanding of robots’ locomotion capabilities for optimal solutions. Traditional approaches that simply classify binary occupancy and understand the planning cost as distances, are unsuitable for legged robots confronting complicated navigation tasks.

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 with locomotion capabilities awareness.

A local motion cost predictor provides cost estimates to a sampling-based planner and a gradient-based optimizer. The planner achieves fast computation time by sampling a fixed set of motions batched on GPU. The gradient-based optimizer refines the raw path by applying cost-gradients to path nodes.

The locomotion cost predictor is trained in simulated environments that contains structured terrains, irregular terrains, and narrow paths.

We compare the path costs from three planners: RRT*, our planner without optimization (raw), and our whole framework with path optimizer (optimized). In all kinds of testing environments, the performance of our raw path planner is competitive with RRT* and our optimized approach can further reduce the path cost means and variances.

Our approach is capable of planning and optimizing paths 3 orders of magnitude faster than RRT* on GPU-enabled hardware, enabling real-time deployment on mobile platforms.

The proposed framework provides smooth and globally optimized solutions when planning on unstructured terrain in mountain, navigating through stairs, avoiding rough terrain, and planning on a narrow path.

We successfully evaluated the framework on the ANYmal C quadrupedal robot in real-world environments for path planning tasks on multiple complex terrains. Our framework navigates the robot to explore along a crowded corridor and move through obstacles based on the elevation map that is continuously constructed by the robot.

Our framework returns navigation solutions with locomotion capabilities awareness of the robot. It prefers flat ground than the low-lying obstacle, while if other ways are blocked, it can also navigate the robot to walk across the obstacle.