Locomotion Strategy Selection for a Hybrid Legged Wheeled Robot

Saudabayev, A., Kungozhin, F., Nurseitov, D., Varol, H.A.,
"Depth Image based Terrain Recognition for Supervisory Control of a Hybrid Quadruped", IEEE International Symposium on Industrial Electronics (ISIE), pp. 1532 – 1537, 2014. IEEE Xplore

In ARMS laboratory, we have developed the Nazarbayev University (NU) Hybrid Quadruped (see Fig. 1). The robot in fact is comprised with two structures – four legged quadruped and four wheeled mobile platform, which does not fall into any of the two types of conventional hybrid robots. The physical separation of the mechanisms is accompanied with the ability of both independent and collaborative actuation for movement.

NU Hybrid Quadruped

Fig. 1. NU Hybrid Quadruped image (top) and CAD model (bottom).

Main idea of the NU quadruped is application of type of the locomotion best suited for specific terrain type, i.e. legs for unstructured and wheels for structured environments. For this, a supervisory control system, which performs a terrain-based locomotion strategy selection, is required. The controller in its turn solely dependent on the terrain recognition capabilities and robust locomotion controllers designed for particular environments. See the robot's control system block diagram in the figure below:

NU Quadruped Control
Fig. 2. NU Hybrid Quadruped control system

In this hieararchical, three-level control, the terrain recognizer is responsible for environment recognition to allow generation of high-level locomotion controller selection. The controllers in their turn reside in the middle layer. They are programmed for robot navigation in specific environment, they consist of sub-controllers and are connected with arrows to indicate possible transitions (see Fig. 3)

Middle layer locomotion controllers
Fig. 3. Middle layer locomotion controllers

The experiments were designed to evaluate and characterize the system. The terrain classification model was created using data collected from five different terrain types - Even, Uneven, Stair Up, Stair Down, Nontraversable (see Fig. 4 for examples). The framework included confidence-based depth image filtering, depth feature extraction, Support Vector Machine classifier model generatino and majority voting filtering. Researchers achieved 97% terrain recognition accuracy for testing dataset.

Terrain Types
Fig. 4. Terrain types (a) Even Terrain, (b) Nontraversable, (c) Stair Down, (d) Stair Up, (e) Uneven Terrain

Final real-time locomotion strategy selection experiment included a route across five terrains (see Fig. 5). The robot successfully navigated through the route with autonomous locomotion switching. No false locomotion was applied by the quadruped. The video of the final experiment:

Funding source: Research project grant funded by the Kazakhstan Ministry of Education and Science