第5届国际机器人智能技术会议成果概览：智能系统与应用前沿

需积分: 9 60 浏览量更新于2024-07-18 收藏 22.49MB PDF 举报

《机器人智能技术与应用5》是一本深入探讨最新机器人智能技术的英文专业书籍，收录在"Advances in Intelligent Systems and Computing"系列第751卷中。该书由Jong-Hwan Kim、HyunMyung、Junmo Kim、Weiliang Xu、Eric T Matson和Jin-Woo Jung以及Han-Lim Choi等专家共同编辑，其内容源自第五届国际机器人智能技术与应用会议的成果。本书的核心主题集中在以下几个方面： 1. **机器人智能技术**：书中详细介绍了机器人领域的智能技术进步，包括计算机智能、软计算（如神经网络、模糊系统、进化计算及不同理论融合）、社会智能以及环境感知（如ambient intelligence）等前沿概念。这些技术为机器人设计提供了坚实的基础，使它们能够更好地适应复杂环境并执行任务。 2. **计算智能与设计方法**：强调了理论研究在推动机器人技术发展中的关键作用，包括算法创新、模型优化以及对人工智能原理的应用，以提升机器人的自主性、学习能力和决策能力。 3. **跨学科融合**：书中涵盖了工程、自然科学、计算机与信息科学、信息技术、经济、商业、电子商务、环保、医疗健康和生命科学等多个领域的交叉应用，展示了如何将智能系统和计算技术应用于实际场景，提高效率和生产力。 4. **社交智能与人机交互**：随着技术的发展，机器人不仅要执行物理任务，还需具备理解人类行为和情感的能力，这涉及社交智能的研究，旨在增强人机之间的自然交互体验。 5. **案例研究与应用实践**：书中还包含了一系列来自国际会议的实证研究和应用案例，展现了机器人智能技术在制造业、服务行业、医疗保健等领域的具体应用，为读者提供了丰富的实践经验参考。《机器人智能技术与应用5》作为一本高清的、最新版本的经典著作，是所有对机器人领域有深入研究兴趣的读者、工程师和研究人员的重要参考资料，帮助他们紧跟行业发展趋势，解决实际问题，并推动智能机器人技术的进一步创新。

Artiﬁcial Intelligence Approach to the Trajectory Generation and Dynamics . . . 5

no related article. The peristalsis scheme is inspired by the biological swallowing

process in human beings by keeping the anatomy of the oesophagus in mind. The

conduit of the robot is manufactured by using a soft elastomer powered by air and

hence, it shows a continuum behaviour [9, 10]. Due to the robot’s complex, and

compliant nature, traditional modelling and control approaches are diﬃcult to apply.

Hence, this paper emphasises more on machine learning based regression techniques

over other conventional methods. The technique can be further extended to imple-

ment control algorithms which utilises powerful regression techniques like Support

Vector Machines (SVM), and Relevance Vector Machines (RVM) [11, 12].

2 Generating Diﬀerent Peristalsis Pattern in the

Swallowing Robot

2.1 Matsuoka Oscillator

A group of neurons or oscillators coupled with each other to produce the desired set

of a rhythmic pattern is known as a Central pattern generator (CPG) neural circuit.

One of the examples of a CPG neural circuit is Matsuoka Oscillator (MO). Matsuoka

introduced the neural oscillator as shown in Fig. 1 which includes two diﬀerent types

of neurons, ﬂexor and extensor neuron where an individual neuron can be represented

by two simultaneous ﬁrst order nonlinear diﬀerential equations [13]. The fundamen-

tal idea behind the working of MO is mutual inhibition of neurons where each neu-

ron is generating a periodic signal to inhibit the other neuron. In Fig. 1, extensor and

ﬂexor neurons are represented by the blue and green ellipse respectively. The vari-

able x

is known as the membrane potential or ﬁring rate of the ith neuron, variable

represents the degree of adaptation or fatigue, c is the external tonic input which

should always be greater than zero, 𝜏 and T represents the rising and adaptation

Fig. 1 Schematic diagram of a two-neuron Matsuoka oscillator

6 D. Bhattacharya et al.

time constant for x

and v

respectively, a

is the synaptic weight which represents

the coupling strength or strength of inhibition between the two neurons, b speciﬁes

the steady-state ﬁring rate, and g

is the sensory feedback weighted by a gain which

is generally the output of a system to which the oscillator is trying to entrain such

that the positive part of the sensory feedback is applied to one neuron and negative

part to the other. The ﬁnal output of the oscillator is represented by y

out

which is

the diﬀerence between the output of the two neurons. Compared to other oscillators,

MO is eﬃcient regarding the order of computation, and the tuning of parameters.

Also, there is no requirement of implementing any post-processing algorithms, and

MO model has a large region of stability as compared to other oscillator. The math-

ematical model of a three-neuron MO is expressed in Eq. 1. The structure of the

connections among the three neuron in the oscillator remains similar to that of the

two neuron oscillator as shown in Fig. 1. The only diﬀerence in case of the former

one as compared to the latter one is the mutual inhibition which is now occurring

between the three neurons instead of two. Due to the complexity in the diagram of

the three neuron MO, only the two neuron counterpart has been illustrated.



𝐱

(t)

𝐱

(t)



=−



𝜏

𝐈

3×3

𝜏

𝐈

3×3

𝟎

3×3

𝐈

3×3





𝐱

(t)

𝐱

(t)





−

𝜏

𝐂

𝐈

3×3



𝐲(t)+



𝜏



111



𝟎

3×1



where,𝐂 =







0 a









𝐱

=[x

,x

]



𝐱

=[v

,v

]

, and x

,x

∈ R

1×3

𝐲 =max



𝐱

, 𝟎



, 𝐲 ∈ R

(1)

2.2 Problem Formulaton for the Parameter Estimation

The problem of designing a MO network is to predict its parameters for generating a

peristaltic wave of desired frequency and amplitude. How the parameters of the oscil-

lator control the wave is not well-deﬁned all the time [14, 15]? A three neuron MO

deﬁned by Eq. 1, has been implemented on MATLAB and the parameter values were

chosen randomly in such a way that they satisfy the oscillator stability constraints.

The estimation of parameters of the three neuron MO falls under the category grey

box modelling as the system of diﬀerential equations representing the oscillator net-

work are known. The parameters govern the shape of the oscillation (amplitude and

frequency), but as the diﬀerential equations are non-linear in nature, so it is hard to

ﬁnd out the particular set of parameter values which represents the desired amplitude

and frequency of oscillation. One way to obtain the parameters is by formulating a

quadratic programming problem on the parameters of the MO and then, solving it

by applying an optimisation algorithm. By using Lagrange’s multiplier in the cost

function, the limit on the size of the parameters can be constrained. The fundamental

concept behind applying such algorithm is to optimise the parameters of the oscil-

lator to return the user-deﬁned shape (amplitude and frequency) of the peristaltic

8 D. Bhattacharya et al.

3 Optical Motion Tracking and System Identiﬁcation

of the Swallowing Robot

3.1 Swallowing Robot’s Motion Capture by Vicon

To capture the deformation of the conduit of the SR, a ﬂat version of the robot was

developed to approximate the original robot. We used optical motion capture system

by installing the Vicon cameras at the several locations of the motion capture lab.

In total, twelve cameras were utilised out of which four were set next to the robot.

An array of 20 hemispherical markers of size 3 mm was glued on the actuator. The

marker array was covering almost one-third of the entire deformation surface. A

spacing of 2.5 cm was maintained between them so that during the actuation the

markers do not interfere with each other’s path. In addition to the small markers,

three big reference markers of diameter 7 mm were placed at three corners of the

plastic frame, so that the location, orientation, and alignment of the robot can be

observed on the TV screen. A tape was used to mask the plastic frame of the robot

as illustrated in Fig. 3a so that the cameras do not capture any insigniﬁcant objects

surrounding the actuator of the robot. The motion capture experiment starts after the

calibration of the cameras and setting the origin of the coordinate system. The centre

of one of the large markers was chosen for the before mentioned purpose. Once the

setup and the calibration have been done, data capturing was started by inﬂating

the chambers of the robot by generating diﬀerent peristaltic trajectories discussed

in the previous sections. Figure 3 shows the images from the experiment conducted

on the ﬂat swallowing robot. The Vicon tracker software provided with the system

can plot the marker’s position data in real time as well as it can be used to generate a

spreadsheet ﬁle where the entire time series data can be stored for further processing.

(a) Flat Swallowing Robot, (b) LCD Screen

Fig. 3 Image of the ﬂat swallowing robot and an array of optical motion capture markers which

are placed on top of it to visualise and record the displacement of the chambers of the robot on a

TV screen

Artiﬁcial Intelligence Approach to the Trajectory Generation and Dynamics . . . 9

3.2 Processing the Captured Time-Series Data

The time-series data recorded from the Vicon tracking experiment needs to be pro-

cessed before using it for the system identiﬁcation of the SR. The ﬁrst step is to

interpolate the missing frames (cells) in the data which the cameras were not able to

capture. Further processing on the data can only be done, once the missing data were

generated. The next step is to remove any particular trend or pattern from the data

which are not any interest to us. Such kinds of aspects are the major cause of some

distortion which can eventually lead to incorrect identiﬁcation of the dynamics of

the robot. By using system identiﬁcation toolbox in the MATLAB, both the process-

ing tasks can be performed satisfactorily. The ﬁnal step is to design a ﬁlter that can

minimise the eﬀect of noise which has been added during the data capture experi-

mentation. The nature of the noise was taken out to be additive, and distributed nor-

mally. A sixth order Butterworth ﬁlter with suitable cut-oﬀ frequency was designed

to remove the ripples from the captured time-series data.

3.3 State Initialisation

The Vicon experiment is an example of data capturing in the Lagrangian frame of

reference. The preliminary tasks before conducting the Vicon experiments are set-

ting the origin and the coordinate axis. For an initial modelling purpose, the second

column of markers as shown in Fig. 4 was chosen from the array. As the movement

of the markers in the x and y direction of the chosen coordinate system was negli-

gible hence, only z-direction displacement was selected. Let





i=1

be the chosen

states deﬁning the displacement of the markers in z-direction and 𝐟 is the unidenti-

ﬁed governing dynamics of the SR respectively expressed by Eqs. (3) and (4). The

movement of the three markers along the z-axis was considered for a preliminary

testing purpose. A training dataset of size 401×3was taken out from the original

processed dataset of size 2923 × 3 (see Sect. 3.1) for modelling the dynamics of the

actuator where the row index denotes the time stamp and column index denotes the

Origin (Center of the

Marker)

()

Fig. 4 State initialisation scheme

剩余562页未读，继续阅读

J-10

粉丝: 18
资源: 483

第5届国际机器人智能技术会议成果概览：智能系统与应用前沿

Robot Motion Planning and Control

- 2016年在线获取 ScienceDirect CAAI Transactions on Intelligence Technology.

robotframework and用法

What is Robot and Robotics?

Q3. Explain Various components of a Robot?

sensible intelligence robot

Explain Various components of a Robot?

<node name="robot_state_publisher" pkg="robot_state_publisher" type="robot_state_publisher" />

robot python

robotframework中robot常用命令

最新资源