Efficient Hybrid Environment Expression for Look-and-Step Behavior of Bipedal Walking

Abstract

The look-and-step behavior of biped robots requires quickly extracting planar regions and obstacles with limited computing resources. To this end, this paper proposes an efficient method representing the environment as a hybrid of feasible planar regions and a heightmap. The feasible planar regions are used for footstep planning, preventing the body from hitting obstacles, and the heightmap is used to calculate foot trajectory to avoid foot collision during the swing process. The planar regions are efficiently extracted by leveraging the organized structure of points for nearest neighbor searches. To ensure safe locomotion, these extracted planar regions exclude areas that could cause the robot's body to collide with the environment. The proposed method completes this perception process in 0.16 s per frame using only a central processing unit, making it suitable for look-and-step behavior of biped robots. Experiments conducted in typical artificial scenarios with BHR-7P and BHR-8P demonstrate its efficiency and safety, validating its effectiveness for the look-and-step behavior of biped robots.

Fig. 1.

Outline of the proposed perception method.

Fig. 2.

Feasible planar region extraction process. (A) Example scene: inclined planes with obstacles. (B) Plane detection result. (C) Polygonization result. (D) Obstacle area removal result.

Fig. 3.

Feasible planar region generation process. (A) Real-world scene. (B) Plan detection result. (C) Polygonization result. (D) Result after removing dangerous areas.

Fig. 4.

Node expansion is performed in general planar regions and feasible planar regions. Candidate footsteps (depicted as red circular areas) are quickly excluded if they lie outside the feasible planar region, which is generated by removing dangerous areas from the general planar region.

Fig. 5.

Construct heightmap. (A) Color image of the actual scene. (B) Corresponding heightmap.

Fig. 6.

(A) Heightmap. (B) Staircase height curve of the barriers during swing process.

Fig. 7.

Plane detection. (A) Quadtree construction. (B) Region growing: The seed node, containing the most points, is identified and merged with its neighboring node that has the most points. The red box represents the seed node, and the green boxes represent its neighbors. The merging process continues until no neighboring nodes can be merged. (C) Result of region growing. (D) Result of plane detection: The final outcome of the plane detection process.

Fig. 8.

Node graph construction. The nodes can be arranged based on their row and column indices and width, which allows for determining the adjacency relationships between the nodes. For example, the nodes adjacent to node A are B, C, D, E, F, G, H, J, L, M, and N.

Fig. 9.

Depiction of BHR-7P. (A) BHR-7P. (B) Abstracted model of BHR-7P.

Fig. 10.

Depiction of BHR-8P. (A) BHR-8P. (B) Abstracted model of BHR-8P.

Fig. 11.

(A) Example scene: stairs. (B) Plane detection result. (C) Polygonization result. (D) Results of footstep planning.

Fig. 12.

BHR-7P climb-up stair experiment.

Fig. 13.

(A) Example scene: steps and a slope with an obstacle. (B) Plan detection result and the planned footsteps before removing dangerous areas. The red arrows indicate several unsafe footsteps that could potentially cause collisions between the robot’s body and the environment. (C) Feasible planar regions after removing dangerous areas. (D) Planned footsteps within feasible planar regions.

Fig. 14.

Traversing through steps and a slope with an obstacle.