ContactNet: Online Multi-Contact Planning for Acyclic Legged Robot Locomotion

TL;DR

This work proposes ContactNet, a fast acyclic contact planner based on a multi-output regression neural network, which is the first time a contact planner is presented that does not exhibit an increasing computational time on irregular terrains with an increasing number of gaps.

Abstract

In legged logomotion, online trajectory optimization techniques generally depend on heuristic-based contact planners in order to have low computation times and achieve high replanning frequencies. In this work, we propose ContactNet, a fast acyclic contact planner based on a multi-output regression neural network. ContactNet ranks discretized stepping regions, allowing to quickly choose the best feasible solution, even in complex environments. The low computation time, in the order of 1 ms, makes possible the execution of the contact planner concurrently with a trajectory optimizer in a Model Predictive Control (MPC) fashion. We demonstrate the effectiveness of the approach in simulation in different complex scenarios with the quadruped robot Solo12.

Figures (6)

• Figure 1: Solo12 robot traversing generated terrain with randomly removed squares of 5x5 $\mathrm{cm}$ dimensions.
• Figure 2: Example of the evaluation on the ContactNet on a terrain composed of stepping stones. Red disks represent some of the actions evaluated by the ContactNet. The others are not shown for image clarity. The network computes the ranking order, according to which the yellow disk is the one which minimizes the cost function \ref{['eq:cost']}. In this example, knowing the terrain map, yellow disk is discarded because it corresponds to an hole in the terrain. Checking iteratively in the ordered output of the ContactNet, the blue disk is chosen since it corresponds to the first action deemed safe.
• Figure 3: Block scheme of the entire locomotion framework. Given the user-defined velocities $\mathbf{v}^\mathrm{usr}$ the actual robot state $\mathbf{x}_\mathrm{c}$, and actual foot locations $\mathbf{p}_\mathrm{f}$ the ContactNet computes in a few milliseconds the gait sequence $\boldsymbol{\delta}$ and touchdown points $\mathbf{p}_\mathrm{f}^\mathrm{d}$ for the three following steps, at a frequency of 3.125 Hz (after each touchdown). Given the sequence as parameter, the Trajectory Optimizer openaccess computes trajectory and tracked by a 1 $\mathrm{kHz}$ centroidal whole-body controller and a joint space impedance controller Grimminger2020. In order to guarantee that the motions are feasible also on the real robot, the torques are saturated to the maximum values that the motor of Solo12 can produce.
• Figure 4: Gait schedule of a walk motion on a stepping stones scenario. White parts indicates moment in which that leg is in swing. The ContactNet finds a completely acyclic gait.
• Figure 5: A comparison between fixed cyclic and optimized acyclic gait selection: Solo12 fails to traverse a narrow stepping stone scenario when ContactNet is only allowed to adapt foot locations. Solo12 traverses the terrain when adapting both footholds and gaits online.