Friday, June 10, 2011

Fast Robots are Hard to Build

Back when I was delusional enough to think I should try to actually win the 2011 Sparkfun AVC, I realized I needed a fast robot. The 2010 winning time worked out to a speed of something like 5mph while the 2011 winning speeds were a little over 10 mph, about what I was aiming for, myself.

I wonder what the winning speed will be in 2012, assuming no changes to the rules? 20mph? 30? more? Several of the robots entered this year were capable of running at more than 30mph. A Traxxas Rustler can peak out at 35mph with a 7-cell battery and a large pinion gear.  The VXL version with LiPo and brushless motor can run 60mph.

But, speed without proper control simply won't get it done.

One of my big goof-ups in building Pokey, the firefighting robot, was that I didn't carefully think about speed, timing, and sensor range. The faster your robot is, the more carefully you need to deal with timing, speed, physics, and sensor range.  Here's what I learned from that failed experiment.

Sensor Range

My AVC robot was originally designed to be speedy (10-20mph), but the 2011 AVC incorporated giant barrel obstacles to detect and avoid.

Imagine dodging boulders on the highway while driving 100mph--in a fog with 10 foot visibility! No way. Sensor range has to be long enough to allow the robot to react to obstacles.

I considered several long range sensor options.  To illustrate the issues: one Sharp IR ranger has a range extending to 150cm.  The MaxBotics LV-EZ1 sonar extends to 6.45m at best. Let's say the robot runs at 8mph or ~358 cm/s. That means the robot will only have 419ms to react to the IR range but a full 1800ms for sonar.

The processor can react within a few microseconds, but... there's the steering servo to consider. It's comparatively slow, moving through 60 degrees in 220 ms.  That takes up a whopping 52% of the entire reaction time for the IR but only 12% for the sonar. The time required for tires to grip and suspension to load before turning rate increases may also be a factor.

And what if you want to double your speed to 16mph?  Or double again to 32?  And from hands-on experimentation, sonar readings may not be entirely reliable at > 6m.

I considered slowing down when an obstacle is detected. Doing so would buy more time for the robot to react and change course. Someday when the robot can actually detect an obstacle, I'll try this. :)

Bottom line, if you want to go fast, you better look pretty far ahead. In addition to 6m sonar, there are some other long range options. There's a Sharp IR sensor, the R316-GP2Y0A710YK, that can reach out to about 5.5 meters.

I also considered machine vision to extend object detection range out past 6m. As you might imagine, the learning curve in the area of machine vision is pretty steep, so that's why I tried a CMUcam1. It was starting to look promising until I ran into other issues.

If I had won the lotto I would've spent $9095 of my winnings on a MESA 10m SwissRanger. :)  Ah well.

Timing

System timing is pretty important too.  At one point, Pokey's infrared ranger sensors updated 20 times a second for wall-following, while steering correction happened at least an order of magnitude faster. The robot would make a steering correction to avoid a wall, but wouldn't wait for the next sensor update.  It'd turn the wheel but nothing would (seem to) happen!  So it'd turn the wheel harder... bad news.

This time around, I carefully planned out the 'schedule' for sensor update and control output to ensure a sane approach to course correction.  I checked the EZ1 sonar, IR Rangers, and IMU sensors at a 20Hz rate and made steering corrections every 100ms. A vision system could be run at 10-20Hz to match other sensor scheduling. Depending on the module, GPS updates at 1-10Hz. Even at 5Hz, that's a long period of time without data so dead reckoning (heading + odometry) has to fill in.

Also worth considering is the update rate versus vehicle speed and sensor range. Let's say the robot detects something at 6 meters. How many more meters will the robot drive until the next ranger update?  At 20Hz and 8mph, it'll be 1/20th of 358cm = 18cm. How many degrees can the steering servo turn in this time?  Only 0.18 degrees in this case.  The robot better not be impatient waiting for something to change.

To the MCU, it's analogous to a person steering a giant battleship.

Physics

Pokey was heavy and took time to change velocity (direction or speed) but was also fast. By the time it detected the need to stop or turn, it overshot its mark considerably.  One can largely ignore momentum in a slow and/or light robot.

As you know by now, Data Bus is at its core an RC truck. Based on testing (read: playing with it) in front of the house, it has all the expected handling limits of any car.  It can't take a corner at straightaway speed.  It takes time to stop and start.  It takes time to change directions even after the steering is turned.  If you turn the wheel too fast at full speed, the front tires can break free and lose traction.

All these physical limitations and nuances have to be considered in the design if the robot is to run reliably and quickly. I took a simple, two-fold approach. First, handling improvements in the form of suspension tuning and better tires, tricks taken from my SCCA Autocross days.

Second, run the car comfortably below its handling limits. Wider turns and slower speeds, in other words. A slightly more sophisticated approach to reach higher average speeds would involve slowing or braking before turns.

A far more sophisticated approach was taken by Project240, limiting lateral acceleration, tracking to different speeds around the course (e.g., arriving at waypoint turns). Quite impressive.

My simple approach worked in that the robot could take corners pretty sharply, entering them around 10-12mph, despite the sensor mast that degraded handling considerably. But the solution was far from maximizing speed and handling potential.

Results
In the end, despite actual navigation issues, the scheduling of sensor, steering, and gps updates worked well, and the vehicle's handling easily supported the meager demands placed on it. So at least something worked out ok.

3 comments:

  1. This is good analysis; something I also think about.

    However, it is my understanding that the winners of the 2011 AVC didn't use sensor data for obstacle avoidance, but relied on having an accurate map and accurate navigation to blindly miss the obstacles and navigate the course.

    One of the top finishers wrote something like: if I have to slow down to miss an obstacle in front of me, I will lose, so why bother.

    In my opinion, that approach is a deficiency of a simple and well defined race course. But, those are the rules we have.

    Ted

    ReplyDelete
  2. Good points!

    Truth be told I originally wrote the article a couple weeks before the competition when I thought the barrels might be moved between heats.

    I guess obstacle detection wasn't super important. And several teams abandoned it due to issues (me included).

    For that matter, GPS was proved unnecessary given the team that ran the record/playback system with gyro+encoders.

    I'm curious to see what the rules morph into next year.

    ReplyDelete
  3. A few followup thoughts. The 2012 winning speed was ~18 mph average, mine ran ~16mph average.

    Once again the top entrants had no obstacle detection and avoidance, instead relying on a combination of accurate position estimation, bumpers, and luck.

    I'm finding (so far) that my sonar won't work at speed. More experiments required.

    As speeds increase above about 30mph, tire ballooning will become an issue for those of us using wheel encoders.

    Pushing handling limits in the corners will, I think, matter more than top speed. More sophisticated approaches for handling may be required to finish at the top.

    Handling will be even more important if the course has more turns or obstacles to dodge.

    I don't think I can dodge the obstacle avoidance bullet much longer.

    Sharp IR rangers may be too slow to be useful at very high speeds.

    We can expect a new venue likely with better GPS reception. That means more receivers will be suitable and more folks may finish (or, at least they'll make it farther)

    ReplyDelete