Real-Time Navigation for Bipedal Robots in Dynamic Environments Octavian A. Donca1 Chayapol Beokhaimook2 and Ayonga Hereid1 Abstract The popularity of mobile robots has been steadily

2025-04-29 0 0 3.63MB 7 页 10玖币

侵权投诉

Real-Time Navigation for Bipedal Robots in Dynamic Environments

Octavian A. Donca1, Chayapol Beokhaimook2, and Ayonga Hereid1

Abstract— The popularity of mobile robots has been steadily

growing, with these robots being increasingly utilized to execute

tasks previously completed by human workers. For bipedal

robots to see this same success, robust autonomous navigation

systems need to be developed that can execute in real-time and

respond to dynamic environments. These systems can be divided

into three stages: perception, planning, and control. A holistic

navigation framework for bipedal robots must successfully

integrate all three components of the autonomous navigation

problem to enable robust real-world navigation. In this paper,

we present a real-time navigation framework for bipedal robots

in dynamic environments. The proposed system addresses all

components of the navigation problem: We introduce a depth-

based perception system for obstacle detection, mapping, and

localization. A two-stage planner is developed to generate

collision-free trajectories robust to unknown and dynamic envi-

ronments. And execute trajectories on the Digit bipedal robot’s

walking gait controller. The navigation framework is validated

through a series of simulation and hardware experiments that

contain unknown environments and dynamic obstacles.

I. INTRODUCTION

Bipedal robots have been a popular ﬁeld of research, due

to the large range of tasks they can be utilized in. The

general humanoid shape of bipedal robots allows them to

be better integrated into a society designed for humans.

However, for bipedal robots to be fully integrated into

society, robust autonomous navigation systems need to be

designed. These systems can generally be divided into three

stages: perception, planning, and control. It is only with the

combination of these three stages into a single navigation

framework in which bipedal robots can truly integrate into

human society.

In this paper, we seek to implement a holistic bipedal

robot navigation framework to enable the exploration of

unknown, dynamic environments. The three components of

autonomous navigation – perception, planning, and control

– must be combined into a single navigation system. The

perception system must be capable of extracting obstacle

and environment structure information. This environment

information must then be used to generate maps of the global

and local environment for planning. A planning framework

must generate global paths and local trajectories that are

robust to unknown environments and dynamic obstacles. Fur-

thermore, planning must respect the kinematic constraints of

the robot while avoiding obstacles to ensure safe and feasible

navigation. Finally, these trajectories must be executed with

*This work was supported in part by the National Science Foundation

under grant FRR-21441568.

1Mechanical and Aerospace Engineering, Ohio State University, Colum-

bus, OH, USA. (donca.2, hereid.1)@osu.edu.

2Ottonomy Inc. chayapol.beokhaimook@gmail.com

a low-level controller that maintains safe and stable walking

gaits. The combination of these capabilities will enable the

safe navigation of bipedal robots in complex environments.

Many works have expanded on the methods of A* [1]–

[5], PRM [6], and RRT [7]–[9] to the unique problems of

bipedal motion planning and footstep planning. However,

many of these works lack several components required for

autonomous navigation systems such as real-time perception,

mapping, and localization processes. Furthermore, only few

works expand further to adapt these bipedal motion planning

methods into more holistic bipedal navigation frameworks.

However, many are still unable to address components re-

quired in holistic autonomous bipedal systems such as on-

robot perception systems [10], localization methods [11],

robustness to dynamic obstacles [12], [13], or validation in

hardware.

We propose a real-time navigation framework for the

Digit robot based on Move Base Flex [14], as shown in

Fig. 1. The framework utilizes two RGB-Depth cameras and

a LiDAR sensor for perception. The environment is mapped

using global and local costmaps to capture large-scale en-

vironment structure and local dynamic obstacles. Odome-

try and localization are calculated using LiDAR Odometry

and Mapping during navigation. We developed a two-stage

planner to generate collision-free paths and obstacle avoiding

trajectories. In particular, a D* Lite global planner capable

of fast re-planning is used to generate high-level paths in

the global costmap. A Timed-Elastic-Band local planner then

follows the global path through optimization-based trajectory

generation that respects kinematic, velocity, acceleration,

and obstacle avoidance constraints. The local trajectory is

executed by generating a sequence of velocity commands

sent to Digit’s walking controller.

The rest of this paper is organized as follows. Section II

introduces the perception, mapping, and localization process.

Next, Section III describes the motion planning methods.

Section IV introduces the simulation and hardware experi-

ments and results. Finally, Section Vconcludes the naviga-

tion framework and provides future work discussion.

II. DIGIT PERCEPTION, MAPPING,AND LOCALIZATION

In this section, we describe the process of building a

real-time map of the environment using Digit’s perception

sensor suite and localizing Digit within that map. The robot

is equipped with two depth cameras (Intel RealSense D430,

placed at the pelvis, with one facing forwards at a downward

angle and one facing rearwards at a downward angle), and

one LiDAR sensor (Velodyne LiDAR VLP-16, placed on top

of Digit’s torso).

arXiv:2210.03280v1 [cs.RO] 7 Oct 2022

Fig. 1. The proposed navigation framework, an architecture built on top of Move Base Flex [14]. Point cloud detections are

used for obstacle segmentation by Random Sample Consensus [15]. Global and local costmaps [16] are generated from the

obstacle segmentations. LiDAR Odometry and Mapping [17] localizes the robot. A D* Lite global planner [18] uses the

global costmap to generate an optimal, collision-free path, which is used by the Timed-Elastic-Band local planner [19] to

generate local obstacle-avoiding trajectories, which are executed through velocity commands sent to Digit’s gait controller.

Fig. 2. Pre-processing and obstacle segmentation results. a)

Original point cloud, b) Filtered cloud, and c) Filtered point

cloud with segmented obstacles in red.

A. Perception

Point Cloud Pre-processing. Before these point clouds can

be used for obstacle segmentation, pre-processing is applied

to obtain a uniform density and remove outlier detections.

First, an average point cloud size reduction of 91.07% is

achieved using a Voxel Grid ﬁlter [20] for downsampling.

Additionally, inaccurate detections outside of the sensors’

accurate range are removed by removing points further

than 2.9 m of the depth cameras. Finally, erroneous points

detected underground due to reﬂections are removed using a

pass-through ﬁlter.

Obstacle Segmentation. After ﬁltering the point cloud from

both depth cameras, the resulting clouds are fused into

a combined point cloud. The Random Sample Consensus

(RANSAC) method proposed in [15] is used to segment

obstacles from the fused point cloud. For a given point cloud,

P, RANSAC randomly samples 3 points to solve a unique

plane model:

ax +by +cz +d= 0,(1)

where a, b, c, d ∈Rare the ﬁtted coefﬁcients, and (x, y, z)∈

R3represents the Cartesian coordinates of a point. Then,

the absolute distance, Di, of each point iin the cloud is

calculated for the ﬁtted model:

Di=

axi+byi+czi+d

√a2+b2+c2

,for i∈ {1,2, .., n},(2)

where n, is the total number of points in the point cloud P.

Points within a given threshold distance, Dthreshold, to the

plane model are labeled as inliers of the model, denoted PI,

PI={pi∈P|Di< Dthreshold}.(3)

This process is repeated iteratively for a speciﬁed number of

iterations, N, determined statistically as:

N= round log (1 −α)

log (1 −(1 −ε)3),(4)

where αis the desired minimum probability of ﬁnding at

least one good plane from P, usually within [0.90 −0.99],

and ε= (1 −u)where uis the probability that any selected

point is an inlier. The resulting plane model is selected as the

one which generates the largest number of inliers with the

smallest standard deviation of distances. From the resulting

RANSAC plane model, inlier points are retained as ground

points, and outlier points are retained as obstacle points.

B. Mapping

To enable real-time planning in a dynamic environment,

we create two 2D layered costmaps presented in [16] to

represent the global and local environments.

1) Global Map: The global map is used to capture and

store macro-scale information of the environment. To enable

this, the global map uses a lower resolution of 0.1 m and

only updates at a rate of 2 Hz. Cells in the global map have

two possible states:

a) Occupied: Any 3D point detections from the percep-

tion sensors are projected onto the 2D plane of the global

map. Any cells that contain projected points are classiﬁed as

occupied and are considered untraversable.

b) Free: Any cell that is not occupied is considered

free and traversable for the robot. Unexplored regions of the

environment are by default considered to be free.

2) Local Map: The local map is used for obstacle avoid-

ance to capture a smaller region local to the robot where

dynamic obstacles may be present. We use a higher resolu-

tion of 0.05 m and update the local map at 10 Hz. The local

map introduces another cell state in addition to occupied and

free cells:

c) Inﬂated: Cells within a certain distance of obstacles

are inﬂated with non-zero cell costs. These non-zero costs are

used to penalize trajectory planning through regions close to

obstacles while not completely prohibiting trajectories from

entering these regions.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Real-TimeNavigationforBipedalRobotsinDynamicEnvironmentsOctavianA.Donca1,ChayapolBeokhaimook2,andAyongaHereid1AbstractThepopularityofmobilerobotshasbeensteadilygrowing,withtheserobotsbeingincreasinglyutilizedtoexecutetaskspreviouslycompletedbyhumanworkers.Forbipedalrobotstoseethissamesuccess,robust...

展开>> 收起<<

Real-Time Navigation for Bipedal Robots in Dynamic Environments Octavian A. Donca1 Chayapol Beokhaimook2 and Ayonga Hereid1 Abstract The popularity of mobile robots has been steadily.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Real-Time Navigation for Bipedal Robots in Dynamic Environments Octavian A. Donca1 Chayapol Beokhaimook2 and Ayonga Hereid1 Abstract The popularity of mobile robots has been steadily

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: