IROS2019国际学术会议论文集 0494_第1页
IROS2019国际学术会议论文集 0494_第2页
IROS2019国际学术会议论文集 0494_第3页
IROS2019国际学术会议论文集 0494_第4页
IROS2019国际学术会议论文集 0494_第5页
已阅读5页,还剩3页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Combining spiking motor primitives with a behaviour based architecture to model locomotion for six legged robots J Camilo Vasquez Tieck1 Jacqueline Rutschke1 Jacques Kaiser1 Martin Schulze1 Timothee Buettner1 Daniel Reichard1 Arne Roennau1 R udiger Dillmann1 2 Abstract Bio inspired robots take advantage of millions of years of evolution to provide interesting and fl exible solutions for issues related to motion and perception Often they have challenging kinematics classical robotics control mechanisms are not always able to take advantage of them A concrete example of this is LAURON V a six legged robot for space exploration inspired by the stick insects The main goals of this work is to combine classical behaviour based control with motor primitives implemented with SNN for motion represen tation We extend a previously presented bio inspired approach to represent hand and arm motion using motor primitives and combine it with a behaviour based architecture to model different locomotion behaviours for a multi legged robot There are four main components First to model the individual leg motions we use two motor primitives implemented with spik ing neural networks for the swing and stance phases Second to control the motor primitives of each leg there are local behaviours corresponding to each phase and corresponding to each activation pattern Third the activation patterns are used to facilitate multi leg coordination and generate different walking behaviours Fourth a high level control interface inte grates control signals from other sources and activates the pat terns We conducted fi ve different experiments to evaluate our approach in a simulated environment using the Neurorobotics Platform NRP The results show that our modelling approach with motor primitives is fl exible enough to represent different types of motions and also highlight the value of the NRP for robotics development I INTRODUCTION Bio inspired robots provide interesting and fl exible solu tions but they often have challenging kinematics and classi cal robotics control mechanisms are not always able to take advantage of them 1 A human hand a musculo skeletal or tendon driven system and multi legged robots are illustrative examples Nature takes advantage of millions of years of evolution to generate new inspiring control strategies but they are diffi cult to implement and replicate 1 A specifi c example of this is LAURON V 2 4 a six legged robot for space exploration missions 5 This envi ronment is not predictable with diffi cult terrain and presents unexpected situations that required a controller to be fl exible and adaptive The current control of LAURON V uses MCA2 6 with a behaviour based control 4 that implements the generation of walking gaits and the posture control The main goals of this work are to combine classical behaviour based control 4 7 with motor primitives implemented 1 FZI Research Center for Information Technology 76131 Karl sruhe Germany tieck rutschke jkaiser schulze timothee buettner dreichar roennau fzi de 2 KarlsruheInstituteofTechnology KIT Germany ruediger dillmann kit edu Fig 1 General view of the closed loop architecture of our motor control hierarchy From right to left the layers have an increasing level of abstraction In motor control we have the individual leg motor control In low level control we have the local behaviours for each leg In high level control we have the activation patterns for multi leg coordination In higher brain areas we have a high level control interface and other motor functions or control sources with spiking neural networks SNN 8 9 for motion representation and to provide a high level control interface to integrate control signals from other sources This work is inspired by the biological concepts of motor primitives 10 as motion building blocks and the way motion is represented as a hierarchy 11 that allows reuse and combination of motions These insights have been successfully transferred and applied to a certain extent in robotics 12 for example with the dynamic movement primitives 13 and the eigen grasps 14 SNN focus on the biological characteristics of neurons and model more closely the way real neurons work Exploring the capabilities of SNN enables research on the learning mechanisms and information representation in the brain Using SNN we can also run the network simulations on neuromorphic hardware 15 In previous work 16 19 we presented brain inspired mechanisms with SNN for motion representation with a hier archical architecture using motor primitives applied to differ ent neurorobotics systems The modelling used to represent the motion of the arm 17 and the hand 18 are combined and extended to model the locomotion A six legged robot is similar to the hand in terms of kinematic structure but with six legs instead of fi ve fi ngers and we added an additional coordination layer for locomotion The main components of our approach are presented in Fig 1 To model the individual motion of the legs we use motor primitives implemented with SNN On top of the motor primitives we added a behaviour based architecture to coordinate the legs By using different control patterns we 2019 IEEE RSJ International Conference on Intelligent Robots and Systems IROS Macau China November 4 8 2019 978 1 7281 4003 2 19 31 00 2019 IEEE4161 can generate different walking gaits and activate the primi tives There is a high level control interface to the patterns to enable external control or input from other networks For the evaluation we present fi ve different experiments in simulation using the Neurorobotics Platform NRP 20 walking forward in circle in zig zag over an obstacle and with a Braitenberg network This evaluation showcases the fl exi bility of our modelling approach for motion representation using motor primitives and highlights the use of the NRP for robotics development II APPROACH Here we present the details of our bio inspired approach see Fig 1 to model and control different locomotion be haviours To test our approach we use LAURON V 2 LAURON V is a robot inspired by stick insects that has six legs with 4 joints per leg and a total of 26 degrees of freedom DoF together with a complete set of different sensors In previous work 16 19 we presented brain inspired mech anisms implemented spiking neural networks SNN 8 9 for motion representation with a hierarchical architecture using motor primitives A detailed view of the architecture with the networks expanded for one leg is presented in Fig 2 In order to move the robot as a whole we need to be able to control the joints of a single leg fi rst To model the individual motion of one leg we use motor primitives implemented with SNN The kinematics of a robot hand are similar to LAURON V but with six legs instead of fi ve fi ngers so the modelling of a leg is similar to that of a fi nger in 18 but it uses more active joints The way primitives are combined and activated is similar to that of the arm in 17 In Sec II A we show how to use motor primitives to generate the swing and stance movements of a single leg and synchronize it using the ground contact signal We implemented a behaviour based network to make smooth transitions between swing and stance phases and also to make sure that both are not active at the same time We added a behaviour based coordination network to represent different walking gaits by combining the motor primitives of the legs The different leg local behaviours and their interplay is explained in Sec II B We want LAURON V not only to move in a straight line but also to move left and right and overcome obstacles we need to be able to combine and change parameters of the mo tor primitives With the control of a single leg we need to co ordinate the six legs in order to generate a fl uid walking gait By using different control patterns we can generate different walking gaits and activate the primitives Leg coordination is achieved by introducing and implementing the Cruse rules 21 and a central pattern generator see Sec II C We also want to be able to move the robot around for meaningful tasks and combine the different walking patterns For this purpose we added a high level control interface to take the desired motion direction as an input and be able to integrate control signals from other sources see Sec II D With this interface we were able to model walking forward in circle in zig zag over an obstacle and with a Braitenberg network A Modelling leg motion with motor primitives Here we present a detailed explanation for the component leg control from Fig 1 Each leg has four joints and the respective kinematic structure is presented in Fig 3a We use trajectory commands to control the legs We extended the model and added bumpers sensors at the end of each leg to detect ground contact in simulation Contact is used to modulate the changes of the phases to inhibit the motion and to adapt the movement The leg movement cycle has two phases a swing phase with no ground contact and a stance phase with ground contact see Fig 3b The joint is used for moving forward and backward and the joint is used for moving down to the ground and lifting the legs see Fig 3a The min and max values are the same for the and joints for both the swing and the stance primitives The swing and stance movements are modelled as motor primitives as presented in 17 19 and are implemented with a SNN Both primitives control all four joints of the leg Each joint has a base trajectory defi ned with an activation function and a minimum and a maximum value For the joint the base trajectory is a sinus shaped function going from 0 to 1 For the joint the base trajectory is a sinus shaped function going from 0 to 1 and back to 0 To walk back wards two primitives are defi ned with inverse trajectories by swapping the minimum and maximum values for the joint The SNN for the swing and stance primitives is presented in Fig 3c As input between 0 and 1 for the neuron popu lation u of the motor primitive the inverted refl ection value between swing or stance is used The population f u rep resents the activation function for the motor primitive The population g f u outputs the motor commands scaled and weighted for the joints The motor commands are forwarded with the output node to the goal out node to decide if the values received from the swing or stance primi tives will be sent to the joints by checking which behaviour is active and inhibiting the values from the other behaviour For turning left and right the step width is changed for the inner legs of the turn for both the swing and the stance primitives Activation values change dynamically to create a smooth transition between phases and to move in different directions B Leg local behaviours Here we present a detailed explanation for the component leg local behaviours from Fig 1 A behaviour based ar chitecture 7 is used to model the leg local behaviours For each leg there is a swing and a stance behaviour see Fig 4a The patterns are the actual leg coordination and there is a corresponding behaviour for each pattern see Fig 2 The principles for behaviour based networks for adaptive control of a bio inspired robot are introduced and explained in 7 We take from 7 the formalization and the defi nition of a single behaviour and the interplay between behaviours and adapted them to work with SNN A behaviour is defi ned 4162 Fig 2 Detailed view of the closed loop architecture of our motor control hierarchy We have expanded the networks in leg control and leg local behaviour for one leg There are motor neurons for each joint that generate the motor commands There are two motor primitives with SNN for each leg for the stance and swing phases To control the motor primitives there are local behaviours corresponding to each phase and corresponding to each activation pattern The different activation patterns are used for multi leg coordination A high level control interface controls the patterns providing an interface to the experiment control and other networks like the Braitenberg a b c Fig 3 Modelling leg motion with motor primitives a Kinematic structure of one leg it has four active joints the full robot is shown in Fig 2 b Main phases of leg motion of stick insects swing and stance adapted from 22 c The SNN for the swing and stance phases as a three tuple B r a F where r is the refl ection or target evaluation function a is the activity function and F is the transfer function Additionally a behaviour also receives sensor input e and the motivation from higher layers The actuator output u of a behaviour is defi ned as F e u The swing and the stance behaviours are complementary so only one behaviour is motivated at a time In Fig 4a we present how the two behaviours are modelled and connected This way of connecting different behaviours not only works for the swing and stance behaviour but also for other types of complementary behaviours We also use the same structure to make sure that only one walking pattern is active at a time The motivation activity and refl ection of swing and stance are always alternating as shown in Fig 4b The swing and stance behaviours can be infl uenced from the outside in two ways by changing the motivation or by changes in the sys tem state represented by the sensor input e Both mechanisms can extend or inhibit the activity of a single behaviour The Cruse CPG and CPG Local behaviours in leg local behaviour see Fig 2 handle actual leg coordination and activate the underlying swing and stance behaviours that ac tivate the motor primitives These behaviours are motivated by the corresponding patterns in multi leg coordination pat terns C Multi leg coordination patterns Here we present a detailed explanation for the component multi leg coordination patterns from Fig 1 The patterns do the actual leg coordination and will determine which leg should swing or stance and accordingly will motivate the corresponding behaviour of the leg The patterns are also implemented as behaviours that are complementary so that only one pattern is active at a time A pattern class takes care of the motivation activity and refl ection for each pattern The pattern class also detects if the robot is in a state of walking or not The robot is in walking state if a direction is given and a pattern is active The motor commands sent to the legs are only published to the actual joints only if the robot is in a walking state We defi ne two leg groups for a tripod gait group 0 with legs 0 3 4 and group 1 with legs 1 2 5 4163 a b Fig 4 Leg local behaviours a Modelling of the swing and stance phases as complementary behaviours adapted from 7 b Com plementary behaviour of swing orange and stance blue From top to bottom the plots show the motivation the activity and the refl ec tion alternating because only one behaviour is allowed to be active see Fig 5b for leg numbering For the tripod gait the swing and the stance phase will alternate for each leg group The Cruse pattern see Fig 5a implements the fi rst three Cruse rules 21 in the following way The swing phase of one leg inhibits the start of the swing phase of the next leg The start of the stance phase excites the start of the swing phase of the next leg The position of the previous leg excites the start of the stance phase This pattern uses the states of the legs to evaluate the rules and switch between phases The CPG pattern see Fig 5b implements tripod walking and synchronizes the activity of both leg groups This pattern uses the refl ections of the underlying swing and stance be haviours of both groups to wait for each other to be fi nished with the swing or stance phases The CPG Local pattern see Fig 5c also implements tripod walking but with no outer synchronization mecha nism This pattern does not consider in which state the other legs are and controls each single leg independently At the beginning a leg group will be motivated to swing and the other to stance and after that each leg will take into account it s own swing and stance refl ection to determine if it will swing or stance next Since no information from the other legs is used after some time the generated pattern will be asynchronous and the generated gait is unstable D Control interface and Braitenberg network Here we present a detailed explanation for the compo nents control interface and other motor functions from Fig 1 We added a high level control interface to integrate control signals from other sources It is possible to use an a b c Fig 5 Multi leg coordination patterns Schematics to represent and visualize the different patterns In a Cruse in b CPG and the leg numbering and in c CPG local Fig 6 Braitenberg vehicle 20 23 implemented using SNN Input neurons 0 1 2 3 orange detect red Neuron 4 is a bias using other colors as input to rotate It has an attractive behaviour when red is detected as neuron 5 blue inhibits the bias and both output neurons green are activated other network a game pad for interactive control or another control module that represents control signals from higher brain areas The control interface provides control signals for each movement turn left turn right walk forward and backward The different experiments presented in next section are implemented by using this interface in the NRP either sending commands from the experiment control state machine or integrating another network To generate a meaningful behaviour we interfaced a SNN that implements a Braitenberg vehicle 23 using this in terface to control the pattern activation layer Although the mechanism is very simple the resulting behaviours of a Braitenberg vehicle appear to be intelligent and goal oriented The sensor input of a Braitenberg vehicle controls the motion directly Different behaviours can be generated depending on how the sensors are connected to the actuators We used a Braitenberg vehicle that spins around until it detects red and then walks straight towards the red color source see

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网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,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论