IROS2019国际学术会议论文集 1955

Abstract Articulated robots such as the humanoid robot often have multiple joints implemented in its limbs It is common for sensor and controller cables to be routed inside of the limb impacting mobility A network protocol with high efficiency low latency fault tolerance is needed to facilitate communication of sensor and controller data over a more compact communications channel In this paper we propose a ring network protocol with fault tolerance in low latency high efficiency communication over a compact channel The performance of the proposed network protocol was verified in simulation and experiment showing that low latency high efficiency and fault tolerance can be achieved over a single communications channel I INTRODUCTION To achieve complicated tasks in a narrow environment articulated robots such as a humanoid robot have been attracting a lot of attention for several decades It is suitable for moving in diverse environments 3 12 It is required to control many components on the articulated robot such as joints When the articulated robot is operating in a real environment there are many cases where the robot contacts with the environment presents fatal problems for the robot In addition a problem may occur in the firmware of each control device making joints unable to move In such a case the system usually gets stuck because it cannot generate whole body motion and in the worst case the robot may fall over and get damaged However a multi legged robot with multiple limbs can use other limbs even if one limb fails By this the robot can continue to achieve its tasks using its remaining movable limbs For example if one hand cannot be operated during a manipulation task the task can be completed by using the other hand Moreover if a failure occurs in the distributed control device that drives the ankle joint in a biped robot it can be recovered by a system reset In addition along with miniaturization of sensors such as cameras and 3D distance sensors multiple external sensors have been mounted at the tip of each limb to measure the state of the working environment To transfer such data low latency is not required as with joint angle and force data In any failures it is necessary to maintain a stable state by performing recovery operations or degeneracy operations It is thus desirable to keep controlling as many joints as possible Therefore it is necessary to be able to reset only an abnormal device rather than the whole limb In this paper we propose a 1 Honda Research Institute Japan Co Ltd FORESTHILLS EASTWING 1F 4 18 11 Minami aoyama Minato ku Tokyo Japan ryusuke ishizaki n f rd honda co jp 2 Honda R D Co Ltd 8 1 Honcho Wako Saitama Japan new protocol with fault tolerance that allows data requiring broadband to transmit on the same communication line while guaranteeing low latency for control data II RELATED WORK There are several types of communication networks used in modern articulated robots Table I shows a comparison of several networks implemented in modern systems The Multipoint Low Voltage Differential Signaling MLVDS used in Valkyrie 1 and the Controller Area Network CAN used in HRP 4C 2 are bus type networks This network is good for saving the number of wire to realize slim joint but not suitable for high speed communication because it is difficult to match the impedance of the wiring On the other hand the ring network uses only two cables but since it is connected point to point between nodes impedance can be easily matched so it is suitable for high speed communication We thus employ a ring network for our robot TABLE I COMPARISON OF NETWORK PROTOCOL ContentsHondaMLVDS CAN EtherCATARCNETResponsive Link SERCOS TopologyRingBusRingRing BusFree p2p Ring Wire saving 2 opt fiber 2 2 8 CAT5 4 twist pair 2 8 8 CAT5e Latency 75u 16node Rate 1Gbps 100Mbps 100Mbps 5Mbps 500Mbps 16Mbps Efficiency 99 2 Any node Any time Master slave Master slave Token Passing Master slave Master slave Fault tolerant CPUReset if hung up INTP Fault tolerant Degeneracy if wire break Detection of abnormal part Impedance can t be match Loopback Ring Bus Impedance can t be match if multiple connection Concerning performance in low latency and high efficiency EtherCAT 6 is used in Robosimian 3 Hydra 4 TALOS 5 Walk Man 14 etc to communicate within low latency However it uses a Master Slave method every node cannot send data without an instruction from the Master module In addition communication timing is considered when some nodes are added On the other hand ARCNET 7 has a Token Passing method which the transmission right is necessary to avoid data collision By this reason it consumes time As a result latency may increase so the system may not be able to be controlled in real time Responsive Link 8 is proposed by Yamazaki et al which has dedicated lines for real time communication to ensure high performance but these dedicated lines use many wires Although SERCOS II 10 A Ring Network Protocol for Articulated Robots Ryusuke Ishizaki1 Takeshi Misumi2 and Takahide Yoshiike1 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 IEEE3882 used in DLR TORO 9 can communicate with low latency the packet header overhead is large since it uses a packet conforming to Ethernet Therefore it is difficult to improve efficiency In addition due to data collisions occur by CSMA CD 11 a network node retransmits data at random times so the latency cannot be guaranteed CHIMP 12 can reboot the power supply modules on each limb and is designed to be aware of fault tolerances However it is desirable to be able to recover on a node by node basis To achieve complicated tasks the protocol must have both low latency and high efficiency and with a fault tolerance function which can be applied to the wire saving ring type network We target as following functions Improvement of protocol performance in low latency and high efficiency A fault tolerance on the articulated robot This paper is organized as follows Section III shows a new network protocol compatible with low latency and high efficiency Section IV presents a fault tolerance function Section V describes implementation and evaluation of new protocols Finally this paper concludes in section VI III NETWORK PROTOCOL In this section we propose a new protocol for ring networks in low latency and high efficiency The core concept is as follows A a communication method which any node can transmit at any time without any data collision B a packet format with low overhead and two types of packet data packet for normal data transmission and interrupt packet that guarantees in hard real time C on the fly processing as packets pass through nodes A Design of Communication As described in the previous chapter the Token Passing method or the Master Slave method have been used in the ring type network Each node cannot transmit data unless there is a Token or a command from the master Thus the latency increases while the efficiency is decreased For example as shown in Fig 1 b since there are lines that do not have Token or data in the Token Passing method the network efficiency is degraded If there is a mechanism that all nodes can transmit at any time without transmission right network efficiency can be improved In addition since each node does not need to wait for the arrival of the transmission right it also leads to a reduction in the latency required for transmission Therefore we propose a communication method that allows packets to flow simultaneously on all lines between network nodes Fig 1 a shows the structure of the network controller NC for realizing the communication method It consists of the following main four blocks Relay block Bypass data externally input Discard the packet sent by its own node Increment HOP value of the packet header Transmit block Send data by command from CPU of own node Packet Manager block Switch data to be output for data input from Relay block and Send block Receive block Interpret whether it is a packet addressed to its own node then receive or discard it Since each network node operates asynchronously there is a possibility to overlap the timing of relaying the data sent by other nodes and the timing of its own sending out Currently both of data collide and then corrupt To prevent this the Packet Manager has a function of sending out the previously inputted data preferentially buffering the other data until the priority packet passes through and sending it after passing Node A Node D Node C Node B a Our Approach Anytime Anynode Relay Transmit CPU Packet Manager Network Controller Receive Network Node Packet from D Packet from C Packet from A Packet from B Node A Node D Node C Node B T Token Free Token Received Transmit to D Receive Token Free b Coventional Method Token Passing Packet from B Input Packet Output Packet Node A Node D Node C Node B T Node A Node D Node C Node B Node A Node D Node C Node B T T Figure 1 New communication method for the ring network Also when inputting at the same time we decided to prioritize the Send part This is to prevent the transmission latency of the own node from being greatly increased when relay data is continuously input Therefore even if all nodes simultaneously transmit data all lines can be efficiently used without data collision B Implementation of two types of packet with low overhead To prevent increasing latency in packet generation and decreasing efficiency due to packet overhead we designed packet structures to be as simple as possible However in the past it was not possible to guarantee real time processing 3883 because of the large communication load so the communication line was separated 8 But it is difficult to reduce the number of wires It is desirable to be able to communicate joint control data requiring real time property and data requiring broadband although it is unnecessary to real time like external sensor data on the same network line Therefore to realize a network in low latency and high efficiency we designed two types of packets data packet for normal data transmission and interrupt packet to trigger other nodes with low latency Due to the size of overhead the communication efficiency can be low Therefore we designed a packet format in which the header trailer CRC are only 32 bits each The structure is shown in Fig 2 The header and the trailer are each divided into four fields The header and trailer of the data packet and interrupt packet are identical except for the last field of the trailer By setting destination ID DID field located in the trailer to 0 xff it becomes a broadcast packet received by all nodes 1 Data Packet DATP Data packet DATP is used for transferring normal data There are four levels of priority in the data packet PRI field located in the trailer of the data packet and transmission reception processing is performed in order of priority Joint control data can be sent with top priority and external sensor data with low priority Header CodeSOPHOPSID Data 1 512byte TrailerCRC EOPFBCDIDPRI Start of Packet Number of Hopping Source ID Free Buffer Counter Destination ID Priority 4 levels End of Packet a Data Packet DATP Header CodeSOPHOPSID TrailerCRC EOPFBCDIDINT Number of Interrupt Factor 8 factors b Interrupt Packet INTP Figure 2 Packet Format 2 Interrupt Packet INTP Even if our network system can communicate with low latency if a destination node takes too much time to interpret packets due to firmware processing of critical data such as synchronization between nodes may be delayed Therefore we introduce an interrupt packet INTP that contains hardware interrupts to control hardware without CPU processing in the destination node There are 8 interrupt factors that can be used by setting the number of factor in the INT field of Trailer As shown Fig 3 a after setting the DID and INT field INTP is sent from the interrupt packet transmission block in NC by the interrupt request from firmware or hardware in source node Upon receiving this INTP in destination node the interrupt packet receive block in NC generates control signal on the interrupt pin of the number stored in the INT field of this INTP Fig 3 b If these interrupt pins are connected to hardware e g CPU interrupt pin or CPU reset circuit in the circuit they can be used to control hardware in destination node directly Transmit Interrupt Packet Source Node Software Interrupt request INT0 7 Command INT0 7 Hardware 8 Transmit INTP Destination Node Receive Interrupt Packet Receive INTP INT0 Control target0 INT1 Control target1 INT7 Control target7 Generated control signal NC NC a Transmit INTP b Receive INTP Figure 3 The Structure of Hardware Interrupt C Data Flow to realize low latency communication In this section the data flow and its characteristics are described In a conventional network packets are sent from the source node and received by the destination After that destination node sends an acknowledgement ACK packet indicating reception completion to the source node Then one data processing is completed However since the number of transactions is two the efficiency is bad and latency will increase because the next packet cannot be sent until ACK comes In this network protocol the transmitted packet is relayed by all nodes regardless of the destination node and it is circulated to the source node and discarded By returning the packet to the source node it can be determined that the Transmit Data Packet CPU INTP Header INTP Trailer INTP CRC DATP Header DATP CRC DATP Trailer DATP DATA D0D1D2D3D4 Input timing for Packet Manager Input Output One DATP One INTP Relay Transmit Transmit Interrupt Packet Packet Manager a Insertion of INTP DATP CRC DATP Trailer DATP DATA D2D3D4 DATP Header DATP DATA D0D1 INTP Header INTP Trailer INTP CRC b Output Data Figure 4 Priority Hanling of Interrupt Packet destination node has received it so that the ACK packet from the destination node is not necessary Therefore since transmission reception can be confirmed in one transaction 3884 latency can be lowered Also if the destination ID of the input packet is its own node the reception process must be performed but if the packet is relayed after judging the availability of reception the latency will increase Therefore by connecting the Relay and Receive blocks in parallel to the data input line and performing the relay processing and the reception processing on the fly the latency in the relay was reduced Interrupt packets are used to realize hard real time such as synchronization between CPUs The transmit interrupt packet unit is independent of the transmit data packet unit and they are processed in parallel Outputs of these units are connected to Packet Manager and Packet Manager controls transmission When the input timing to Packet Manager is simultaneous an interrupt packet is sent with priority However while a DATP is being relayed if a CPU of its own node requests INTP transmission the INTP will be kept waiting until the DATP passes to prevent collision If the size of the DATP is large it takes time to transmit the INTP so there is a possibility that the real time property may be impaired Therefore as shown in Fig 4 a Packet Manager has a function to insert an INTP by breaking a data packet even during transmitting DATP Fig 4 b shows the output data from Packet Manager at that time As a result low latency transmission of interrupt packets was realized If a node kept transmitting a DATP other nodes cannot transmit it For example when there is a node that keeps flowing data of camera it is implied that the other node cannot transmit DATP such as joint control data that should be transmitted in real time By this reason we also set a constraint that the next packet cannot be sent until the packet returns to own node for each packet In addition we set the maximum size of DATP to 512 bytes Fig 5 shows an example of data flow when Node A sends a DATP to Node C Source ID SID A Destination ID DID C When Node A sends it Node B can relay it because Node A Node D Node C Node B SID A DID C Packet Relay Relay Receive Discard Completion Relay SID A DID C Packet Transmit SID A DID C Packet On the fly 1 2 3 4 5 Figure 5 Data Flow of Proposed Network the node know that the packet was sent by another node by the SID value of the packet header After that it reaches Node C as the transmission destination Since Node C knows that it is a packet addressed to itself by DID it performs reception processing At the same time since the SID is not its own node the packet is relayed The packet relayed by Node C is similarly relayed by Node D and then returns to Node A When Node A confirms that the source is its own node the packet is discarded By this data flow the maximum latency is 75 us even in the state where the communication load is the highest such as 16 nodes sending 512 bytes of data packet at the same time Therefore it can be said that it is a network protocol that can realize low latency IV FAULT TOLERANCE When the articulated robot performs work remotely it is necessary to continue to move without getting stuck even if some failure occurs in a part of the system For example when a firmware error is occurred on a node the system should be recovered by resetting only the CPU of abnormal node to prevent entire system is shutdown It is also desirable that degenerate operation can be performed while allowing the abnormal state even when a hardware error such as a disconnection occurs In Fig 5 it is assumed that Node A is a Main Controller Node to generate all joint commands Node B C and D are Motor Controller Nodes which drive joints If Node A can detect the disconnection between Node C and D Node A know that sensor feedback data from B and C cannot be acquired So it is conceivable to change to feedforward control not using that data Node A generates degenerate movement which is used only the joint which is driven by Node B C However in a ring net