Unexplained intermittent CAN / 2CAN Jaguar problems at GSR

[John Heden]

Greetings All,

We started this thread after experiencing a number of catastrophic CAN startup problems at the GSR regional and while we were hopeful that the V29 update would provide some relief on this issue, we continued to see this problem occasionally at the Hartford regional this past weekend. While our Drive team was instructed to monitor the Driver Station diagnostic tab for streaming CAN errors, they became complacent on Friday after never encountering the issue on Thursday. A Driverstation side CRIO Reboot did NOT recover the situation after the match began and we sat idle during the entire match (our Alliance won without our participation).

We experienced this problem again on Saturday while setting up before opening ceremonies (we had the first match) and again the system did NOT recover with a warm Driver station reboot on the first attempt and required either a third reboot attempt or possibly an actual robot power cycle to recover. This early morning triple failure scared us a bit but we NEVER experienced this catastrophic CAN failure during our later matches. The drive team did occasionally see a few startup CAN errors that concerned (panicked) them but did not see the catastrophic scrolling CAN error behavior.

We saw this failure occur during GSR at frequency of about 1 out of 6 matches ONLY while we were actually on the playing field and never while tethered to the robot with approximately the same statistic at Hartford. There was an observation that was seen once on the practice field making us curious as to whether the use of the radio was some how a catalyzing factor. Our radio was physically touching the 2CAN so we decided to try to give them some space (inverse square law). We couldn't try this with the radio in the pits but we proceeded to do some repetitive tethered power on/off tests trying out different power up sequences (relative to laptop) to try to reproduce this. We must have done this 20 or 30 times and NEVER saw a single CAN transaction failure and certainly not the continuous scrolling catastrophic CAN failure signature. We thought this was a radio ONLY failure but we did eventually experience this once while we were tethered in the pits. A quick power cycle and the problem went away! I wish we had tried a soft reboot as an experiment but our robot was being queued and our goal at that point was recovery rather than experimentation.

I believe we have some type of CAN/2CAN/CRIO/WPILib startup race condition that occasionally prevents some type of low level initialization causing the complete loss of the CAN bus. The manifestation we see is as if we simply pulled the CAN cable out of the 2CAN preventing any successful transaction to any CAN device. I believe use of the radio somehow amplifies this window of opportunity for failure given our ratio of match failures to pit failures and given we power up much more often while tethered in the pits than during actual matches. We had little working radio based experience prior to arriving at GSR due to the late availability of the physical robot for software testing. This radio testing and its influence on CAN failures will be a priority when we get our robot back. My apologies that some of this data is so soft but we were unable to find any hard correlation or anything definitive other than an occasional complete startup failure that always recovers on a power cycle mostly ruling out cabling issues. This failure occurs BEFORE being enabled essentially ruling out any real voltage drop or current/noise problems. If we startup successfully, we do run successfully. In fact, we have performed number of tests where we pull the breakers out of the Jags and even the 2CAN. This causes CAN errors to be reported but the system nicely recovers within a couple of seconds after we plug the breakers back in. We use the default voltage mode so others who have a more complex initialization or control scheme may not recover so easily.

There was also some anecdotal data coming from others at Hartford (other teams and even some of the Harford technical field folks) that believe the serial CAN interface is more robust than the 2CAN and recommended we switch away from the 2CAN. While this startup problem is catastrophic, it feels like some type of simple initialization glitch that is solvable. The CAN & 2CAN approach is a nice technology with perhaps this single gremlin to be exorcised. We'll try to diagnose this further when our bot comes home but unless we can convince our team that this is behind us, we may be forced to return to the simpler ways of PWMs....

Cheers and thanks,

John

1) CAN Wiring is correct with proper termination.
2) All components are ground isolated from the frame and electrical wiring has no shorts or ground faults.
3) We run the CRIO connection directly to the radio and connect the radio directly to the 2CAN rather than passing all CRIO traffic through the 2CAN.
4) We used all Tan JAGS.
5) We Programmed with C++.
6) Used Voltage Mode only.
7) 3 Jags with optical encoders, 1 of these has a single limit switch as well
2 Jags each with 2 limit switches, no encoder

8) I should also add that our software launches a separate dashboard thread at the end of the constructor AFTER the Jags and other robot objects are created with this data (encoder values, currents, voltages, etc) being read for display and capture by our custom dashboard. This explains why we see a continuous never ending data stream while others, I suspect, may see a number of errors during startup that stop but resume once they are enabled and autonomous & control Jaguar transactions begin.

[Hugh Meyer]

John,

I may be misunderstanding your comment # 3. Please clarify if my comments don't make sense.

*****
3) We run the CRIO connection directly to the radio and connect the radio directly to the 2CAN rather than passing all CRIO traffic through the 2CAN.
*****

I thought the 2CAN was to be connected to the CRIO on port # 2. Since port # 2 is on a different network the traffic is isolated from the robot communication traffic on the wireless. That is how ours is wired and we just completed 2 regional events without any control issues. We had other issues, just not control ones...

It seems adding that additional load through the radio switch and robot communication network could indeed cause problems.

-Hugh

[taichichuan]

Originally Posted by Hugh Meyer

John,

I may be misunderstanding your comment # 3. Please clarify if my comments don't make sense.

*****
3) We run the CRIO connection directly to the radio and connect the radio directly to the 2CAN rather than passing all CRIO traffic through the 2CAN.
*****

I thought the 2CAN was to be connected to the CRIO on port # 2. Since port # 2 is on a different network the traffic is isolated from the robot communication traffic on the wireless. That is how ours is wired and we just completed 2 regional events without any control issues. We had other issues, just not control ones...

It seems adding that additional load through the radio switch and robot communication network could indeed cause problems.

-Hugh

Routing CAN traffic through the radio could certainly cause problems. The radio's ports have a limited amount of buffer space before the 802.3x congestion control messages start flying around on the net. I'm not sure what the 2CAN would do if it suddenly started getting a lot of source quench message traffic. Certainly, packets would start getting lost and that would be bad on a half-duplex style network like CAN (at least in the way it's implemented for FIRST). So, it's probably better to wire the 2CAN to port 1 on the cRIO and wire the radio on the other port of the 2CAN.

My $.02,

Mike

[John Heden]

Originally Posted by taichichuan

Routing CAN traffic through the radio could certainly cause problems. The radio's ports have a limited amount of buffer space before the 802.3x congestion control messages start flying around on the net. I'm not sure what the 2CAN would do if it suddenly started getting a lot of source quench message traffic. Certainly, packets would start getting lost and that would be bad on a half-duplex style network like CAN (at least in the way it's implemented for FIRST). So, it's probably better to wire the 2CAN to port 1 on the cRIO and wire the radio on the other port of the 2CAN.

My $.02,

Mike

Mike,

Thanks for your thoughts. We initially had the 2CAN device on the CRIO port 1 connector which is where we first started to see our problems. The buffer capacity of the 2CAN vs. the radio was unclear but our intuition gave the radio the advantage here. I believe our next step will be to simply move to port 2 of the CRIO and see whether that helps things.

Thanks,

John

[nuttle]

Do you know if your 2CAN has had the firmware updated to version 2.5 or not? We have a regional this coming week/end and will try out v29 on the cRIO. We've been using serial, but it is easy enough to switch back and forth that we might try the 2CAN again, at least for practice matches. We have not yet had a chance to try either of these updates.

[John Heden]

Originally Posted by nuttle

Do you know if your 2CAN has had the firmware updated to version 2.5 or not? We have a regional this coming week/end and will try out v29 on the cRIO. We've been using serial, but it is easy enough to switch back and forth that we might try the 2CAN again, at least for practice matches. We have not yet had a chance to try either of these updates.

Yes! We updated to V29 and had V2.5. We started to try a switch to serial but ran into some difficulties trying to craft a proper RS232 to Black JAG cable during the competition as well as finding a RS232 to USB converter. There were a number of folks at Hartford who were suspicious of the 2CAN and suggested that a RS232 conversion might solve our intermittent problem. We remain optimistic that a robust 2CAN solution is possible.

[jtdowney]

Originally Posted by taichichuan

Routing CAN traffic through the radio could certainly cause problems. The radio's ports have a limited amount of buffer space before the 802.3x congestion control messages start flying around on the net. I'm not sure what the 2CAN would do if it suddenly started getting a lot of source quench message traffic. Certainly, packets would start getting lost and that would be bad on a half-duplex style network like CAN (at least in the way it's implemented for FIRST). So, it's probably better to wire the 2CAN to port 1 on the cRIO and wire the radio on the other port of the 2CAN.

My $.02,

Mike

R50A states "The DAP-1522 radio is connected to the cRIO-FRC Ethernet port 1 (either directly or via a CAT5 Ethernet pigtail)." We took this to mean that under no circumstance can an active device (2CAN) sit between the DAP-1522 and the cRIO. That leaves two choices, connect the 2CAN to the DAP-1522 or connect the 2CAN to Ethernet port 2 on the cRIO.

My teams robot (programmed with Java using IterativeRobot) has gone through one event with our 2CAN plugged into our DAP-1522 and have had no CAN related trouble. We were running cRIO v28 (v29 wasn't out at the time) and 2CAN firmware v2.5 with the SVN rev 66 plugin on the cRIO. We have 6 black jaguars on the CAN bus with no sensor inputs or limit switches.

Perhaps we were very fortunate during our regional but we have not had any serious CAN issues (knock on wood) since build. All of our trouble then could be traced back to poorly made cables when we did have problems.

I am hoping our luck caries us through championship.

[mjcoss]

We continue to be plagued with timeouts on the CAN bus. And yes, I've checked the termination, and it all looks good. We are running V29 and I don't have the numbers for the plugin or the 2CAN firmware.

One thing that I have seen which is causing no end of issues is that if you get a timeout on the messages, the API is return no indication of the failure. So, for example, if the GetForwardLimitOK() function is called, and times out, you get back false. There is no way to know that that has happened and if you are making decisions based on these results... We have an encoder on our lift mechanism. To zero the encoder, we drive to the bottom limit switch, and when we get there, we set the encoder to 0. This works fine until we lose the message due to timeout. From that point on the lift is offset by where ever the timeout occurred. There really needs to be a way within the API to detect that the transaction timed out.

Of course, the best answer would be that we don't have any timeouts.

Another observation related to timeouts is that we have an on board compressor and if during initial startup the pressure sensor indicates that the compressor should run and starts the compressor immediately, we get a number of timeout messages.

All in all, I'm really regretting the decision to use the CAN bus. And for the most part all of the features that I really wanted to use, that were provided by the CAN bus, proved to be unusable.

[jhersh]

Originally Posted by mjcoss

One thing that I have seen which is causing no end of issues is that if you get a timeout on the messages, the API is return no indication of the failure. So, for example, if the GetForwardLimitOK() function is called, and times out, you get back false. There is no way to know that that has happened and if you are making decisions based on these results... We have an encoder on our lift mechanism. To zero the encoder, we drive to the bottom limit switch, and when we get there, we set the encoder to 0. This works fine until we lose the message due to timeout. From that point on the lift is offset by where ever the timeout occurred. There really needs to be a way within the API to detect that the transaction timed out.

Yes... that's a definite shortcoming of the C++ API. Both the LabVIEW and Java CANJaguar APIs have a mechanism for the user's program to know that there has been an error. I intended to remedy that this season, but didn't have the time. It will be fixed next year.

Originally Posted by mjcoss

All in all, I'm really regretting the decision to use the CAN bus. And for the most part all of the features that I really wanted to use, that were provided by the CAN bus, proved to be unusable.

I'm sorry to hear that you aren't happy with the CAN implementation. What features did you want to use that you aren't using? What made them unusable for you?

[mjcoss]

Here's what we were planning:

1) We are using a two speed transmission from AndyMark. It has uses 2 motors, and has a single optical encoder built into the the gearbox (on the output shaft). I had originally intended to use speed & position control on the drive subsystem. Position control for autonomous, and speed control during teleop. We first tried just using a y splitter to feed both Jaguars, and then tried a master/slave arrangement. While we got it to work, it seemed unstable. I'm sure that with more time, we might have been able to get it to work reliably but there really wasn't enough time.

2) Our lifter mechanism had the same problem. We had intended to use position control to put the arm at preset heights. Again, we had two banebot motors driving a custom gearbox with an encoder of the output shaft. The position controls would work for some positions, but would oscillate and not find the set point at others. I tried tuning it several times and couldn't get something that was 100% stable. Again we tried a Y split, and a master/slave configuration. We ended up moving the encoder to the CRIO and ran the PID loop on the CRIO. This gave me better control on the loop, and was immediately stable.

3) We have limit switches directly connected to the Jaguars, for the lifter mechanism, and the claw. The bottom one on the lift mechanism was used to home the encoder. We drop to the bottom limit switch and set the encoder count to 0. This works fine unless we timeout on the get limit message. If that message times out, the API returns at limit, and we stop. The premature stop of the motor made zero be in the wrong place, throwing all subsequent offset off by upwards of 6 inches. Apparently, in Java you can catch the exception, but I didn't see anything in the C++ code to allow the caller to know that the results of the call was bogus. I've modified my version of the limit switch calls to return an error status so that I can decide what to do if an error occurs.

Now add into the mix that, under some unknown set of circumstances, we get a flood of CAN bus timeout messages making the robot completely unusable. Given all that, one has to wonder if using the CAN bus is a good idea. In addition, we get some number of intermittent timeouts, which while not fatal causes weird behavior because the code can't determine that a timeout has occurred since the caller is returned a valid return value even if the call times out.

I have updated the firmware to the latest versions. I've tried two different bus terminators. And new cables.

One item that we've seen that has lead me to point a finger at the 2CAN is that on our robot we have an onboard compressor. We have noticed that if during autonomous if the air pressure in the system is low and the compressor needs to run immediately, we get the flood of CAN bus messages. It is as if the run all 4 CIM motors, plus rehome the lifter, plus drop the arm, plus run the compressor is a trigger for the failure My theory is that the voltage drop is causing a failure in the 2CAN. Of course, it could be something completely unrelated.

We are going to St. Louis and I'm seriously debating whether we should rewire the bot to use PWM cables.

[drakesword]

Originally Posted by mjcoss

Here's what we were planning:

One item that we've seen that has lead me to point a finger at the 2CAN is that on our robot we have an onboard compressor. We have noticed that if during autonomous if the air pressure in the system is low and the compressor needs to run immediately, we get the flood of CAN bus messages. It is as if the run all 4 CIM motors, plus rehome the lifter, plus drop the arm, plus run the compressor is a trigger for the failure My theory is that the voltage drop is causing a failure in the 2CAN. Of course, it could be something completely unrelated.

We are going to St. Louis and I'm seriously debating whether we should rewire the bot to use PWM cables.

Its my understanding that the 2CAN wont drop out till about 6 volts. Regardless are you wiring it to the termination blocks or to the 12 regulated supply or the 24 volt regulated supply?

We had ours wired on the 24v line parallel to the solenoid bumper. We had dropouts fairly often but when it happened we saw the PHY lights go out on the connector. We switched the 2CAN to port 2 and had another dropout some time later. We then swapped out our radios (we had issues) and never had a dropout since. We also developed a pre-match procedure where one of the drivers logged into the 2can and checked to see that all the jags were online. Once we started and only one jag showed up. Quick power cycle and all of them were back online.

I'm going to write a quick java applet that will display the 2CAN status information without the need to open a browser (aka have it pop up with the driver station on load)

[mjcoss]

Off hand I don't know for certain how the 2CAN is connected (I'm more of a software guy), but I believe that it powered via the termination blocks. Should it be moved?

Just as an aside, how are people connecting the 2CAN? We have the CRIO connected to the DLINK via the CRIO's port 1 and the 2CAN via the CRIO's port 2. The second connector on the 2CAN is unused. We had though of adding a camera to the robot and I was planning on plugging the camera into the unused port of the 2CAN.

However, when I first wired it up, I had the 2CAN connected to port 1 of the CRIO and the DLINK was connected to the second port of the 2CAN, but it seemed to me that in this configuration, the 2CAN was going to have a lot more work to do that was unnecessary, as it would have to bridge all the traffic to and from the CRIO and driver station.

[drakesword]

Preliminary release for the app is available. Go to this post

[rrossbach]

Our robot has 8 Jags (4 black, 4 grey) running off of 2CAN, and we do a fair amount of runtime data logging on the cRIO for post-match analysis. We see occasional timeouts (approx every 20sec or so) but catch them so it has never really caused a control problem.

Interestingly, even when seeing timeouts we have never seen any CAN errors reported on the 2CAN webdash - assuming that the webdash is working correctly, that leads us to suspect that the occasional timeouts not caused by any wiring/comms issues with the CAN bus itself, but rather something else going on in the software stack within the cRIO.

Once nationals are over we're planning to test a variety of scenarios (running Jags off the 2CAN both with and without the cRIO, requesting tons of status info from the Jags continuously, etc) to try to collect very specific data on what scenarios provoke the timeouts. It's a great learning exercise for the students, and hopefully the data will help narrow down exactly what's going on. Can't promise exactly when we'll dig into this but we plan to post the "scenario design" we come up with and seek input from the community here.

- Ron
Team #2607 controls mentor

[mjcoss]

So I finally got a chance to run some tests on our second bot that was made in parallel with our competition bot. It is not exactly the same but we were seeing similar CAN bus timeouts.

I replace the bus termination with a new one using a 120 ohm resistor (our old ones were 100 ohm), and we move the 2CAN power supply to the 24volt regulated power off the power distribution block (I think that what the electrical guy said0. Over a 10 minute session, cycling thru teleop, autonomous, and disabled states, we saw only 2 timeout messages.

Unfortunately, I can't recreate the one "known" failure case that we have seen on our competition bot. That is, if in autonomous, we need to run the compressor at the start of autonomous mode, we get a complete CAN bus failure. The problem is that our second bot was partial disassembled, and so not all of the components are running/attached.

But it seems like a good sign, although why it would loose even one is a bit of a mystery. Maybe if we increase the wait time a small step to see if the message was just delayed or was really lost.