Path Planning With Local Motion Estimations

Abstract

We introduce a novel approach to long-range path planning that relies on a learned model to predict the outcome of local motions using possibly partial knowledge. The model is trained from a dataset of trajectories acquired in a self-supervised way. Sampling-based path planners use this component to evaluate edges to be added to the planning tree. We illustrate the application of this pipeline with two robots: a complex, simulated, quadruped robot (ANYmal) moving on rough terrains; and a simple, real, differential-drive robot (Mighty Thymio), whose geometry is assumed unknown, moving among obstacles. We quantitatively evaluate the model performance in predicting the outcome of short moves and long-range paths; finally, we show that planning results in reasonable paths. Index Terms-Motion and path planning, deep learning in robotics and automation, probability and statistical methods.

Real-time Optimal Navigation Planning Using Learned Motion Costs

“…In addition, regressing the robot motion outcomes gathered from simulation helps to formulate the planning cost function in a more convenient and intuitive way [6]. A CNN trained with simulated data can outperform feature-based approaches in traversability classification [3] and can be used in a framework that estimates multiple local motion descriptors in continuous configuration space to construct the cost function for path planning using RRT* and Stable Sparse RRT [1]. Nevertheless, it requires a long planning time to successively predict a large number of sampled motions for a smooth and feasible path using a deep neural network.…”

“…We reference the model in [1] to develop the local motion cost predictor: Each robot state q i in the configuration space Q contains the robot's center position x i , y i and the heading direction ψ i in the world frame:…”

“…With the intensive research on legged system control in recent years, quadrupedal robots are presenting increasingly strong capabilities when operating on complex terrains. The robot may have various motion outcomes and undertake different risks when interacting with diverse environments [1] and is expected to move along a smooth and reasonable path with minimum travel expense and failure probability provided the local information. Furthermore, their locomotion capabilities strongly depend on the applied locomotion controllers.…”

“…Local terrain characteristics are directly used to evaluate footprint traversability and construct the cost function with a group of hand-tuned formulas in [2], which may require strong specialty and extensive experience on specific robots when the terrain complexity increases or more cost factors such as energy and time consumption are introduced. [3] adopts a deep neural network that learns local motion cost estimates from simulation, and [1] combines the estimator with variants of RRT for global path planning. However, the pure samplingbased planning frameworks require frequent queries of the Fig.…”

“…We present a navigation framework that supports rapid global path planning and optimization on complex terrains and test it on an ANYmal C [4], a quadrupedal robot with high mobility. We first train an accurate local motion cost predictor in simulation in a similar approach [1], which estimates several motion attributes based on a local height scan and locomotion commands. To search for a globally optimized path in a short time, we rapidly build a roadmap by batch-processing of predefined local motions on a GPU which we use to obtain a raw path using the A* algorithm.…”