Mike and Dave's DARPA Blog

REALbasic Developer Interview

By popular demand, here is a PDF copy of an interview we did for REALbasic Developer magazine.

Spectators

"The hard work by your team clearly showed during the static and dynamic site evaluations, but the competition among the teams was tough. After evaluating the results of all site visits, I regret to inform you Calculated RISC will not be invited to the NQE."

- DARPA, Aug 23, 2005

So there you have it. We were not invited to take part in the National Qualifying Event, but we have been invited to attend as spectators. We'll have to think about that one.

-Dave

Hard work pays off

Monday is a bit of a blur--Mike and I stayed up until 2am working on obstacle detection and avoidance; basically trying to cover all cases we could think of, with the intention of testing each one in the morning. We each slept for literally an hour, and hit the road at 4am. Stopped at WAL-MART for a pack of batteries and some graham crackers, chugging coffee and watching the sun rise over Lake Michigan on the way north. We unloaded the vehicle and had it setup in record time, just as we were getting enough light to start testing. Things were finally starting to look good!

It only took a few trial runs to determine which methods of dodging were working better than others, so we concentrated on tweaking those. Quick explanation: the method we use to detect obstacles is far more capable and correspondingly more complex than what we used at the first site visit. Using only two cameras we can get a very good estimation of the roughness of the ground and accurate heights for objects in the path of the vehicle. We have a vertical resolution of a few inches and can place the objects accurately to within a foot or so of the front of the vehicle. Once we have this 3D representation of the world the driving logic needs to decide how best to navigate it, which brings us back to our trial runs.

The DARPA folks showed up at 8:30, just as we were finishing a successful run(!) We had our 200m course staked out and ready to go. We chatted a bit and explained what bits and pieces had changed since last time, and why we thought they were improvements. They seemed impressed that we had chosen fairly difficult offroad terrain, and wished us luck.

The first run was underwhelming... we got caught in a state wherein the vehicle wouldn't move--conflicting information about obstacles and the corridor edge scared it a little too much. We did run most of the course and dodged one obstacle before it happened, so we got credit for a partial run. On the positive side, the judges did get to see Mike and I troubleshoot and the results of the fix were obvious on the next run.

The second run was better. We hit the first obstacle and dodged the second, which was placed in the middle of a very tight turn at a waypoint. We drove up to it, slowed down, and verrry carefully crept our way around, coming within inches of the corridor flag and the trashcan. It was beautiful :) The judges commented on how well everything worked at that particular waypoint. It continued on to the end of the course with no problems.

Run three wasn't so great... missed the first obstacle, but ran wide out of the course for a little ways and hit a flag on the way back in, then proceeded to flatten the second can and get stuck in the same state as the first run. A real bummer, apparently due to the GPS losing the fix temporarily.

Instead of pushing the vehicle back to the start line, I thought it would be a good idea to demonstrate the capability to run at high speed. We kicked up a good dust cloud on the way back and hopefully proved that we could indeed go pretty quick when we need to. Dodging obstacles at high speed is still difficult for us, but we're ready to tackle that next.

The fourth run is a free-form run up to 1000m, but we hadn't had time to collect GPS points or stake out a longer path. We decided to run the same 200m course we had already used, but hopefully get in another perfect run to offset the disappointing first attempt. We drove the course in reverse and continued up a relatively steep hill to show off a little bit (all unofficial, uncounted driving time), did a zero-degree turn at the top, and drove back down the hill to our starting point. This one was just about perfect. We stayed well clear of the obstacles and ran straight to the end with no incidents. The timing was perfect; my laptop battery died just as the vehicle came to a stop at the final waypoint. We talked to the officials for quite a while after the final run, while they inspected the vehicle and quizzed us about the vision system and the mechanicals. I think we provided good answers for everything.

No we are back to playing the waiting game... DARPA will notify the nine alternate teams of their status on August 22, which is only five days from when I'm posting this. Neither Mike nor I feel that we have any idea where we stand at this point--but we are both looking forward to hearing a positive response next week! It would be incredibly satisfying to make it to the NQE and we're both prepared for the ride.

-Dave

NOTE: Pictures on the way...

UPDATE: Pictures here

On the Nature of Site Visit Preparation

Sorry for not updating from Appleton over the weeked, but there was literally no time to spare outside of vehicle preparation and testing. We were at the field all day Saturday and into Sunday morning (~1:30am) tweaking the drive system, giving the vehicle mechanics a final once-over, and squashing a few remaining driving logic bugs. Basically making sure we could focus on the vision and obstacle avoidance all day Sunday.

We were back at the field Sunday morning--with less than 24 hours until the official site visit--when Murphy's Law decided that it hadn't seen us recently and that it was time to pay a visit. Part of the daily setup process involves checking connections and verifying that the serial ports are all talking to the correct components, and Sunday morning we found that two of the soldered wire connections on the compass had snapped off at the serial port. Now we were dealing with a 30 minute trip to Radio Shack, shooting at least an hour of daylight that we needed for testing. Colin, our videographer, drove me to Green Bay to pick up a soldering iron and a new serial connector.

Good thing too, because by the time we fixed the compass and had all the computers and sensors wired up Radio Shack was closed. Then we would have been in really deep trouble when we ran into the next showstopper, which shouldn't come as a surprise to anyone: blown digital potentiometers on the Whyachi speed controllers. This time it was much more serious because we also lost the PIC that controls the forward/reverse signal, so the circuit was completely useless; we couldn't just drop in another Xicor chip to fix it like we had in the past. We went into disaster recovery mode and quickly rebuilt the servo-driven analog potentiometers. This meant reprogramming the Brainstem controller board and reconfiguring the Sevcon motor controllers, which on a good day takes at least an hour. Factor in the time spent dealing with finicky invertors and failing laptop batteries and you end up with very little daylight left.

All said and done we didn't get nearly as much testing time as we had hoped. Monday morning was clearly going to be crunch time.

-Dave

Apple = progress

Wow, it looks like an Apple store exploded in my living room! We just received a bulk shipment of Apple demo units: four 1.42GHz Mac minis and a dual 2.3GHz Xserve. The vision code we have been developing is fully network-aware thanks to Mike's work on the CarNet protocol, which we developed from the ground up to suit our particular needs. Thanks to network auto discovery adding another processing node is as easy as plugging another Mac into the network and launching the vision node application. The head node that grabs the images from the iSight cameras will send out the left/right stereo pairs, divvying up the work between each processing node according to how many levels of disparity we want returned in the depth map. A single node is currently processing approximately one frame per second, which includes all driving logic decision making and drawing the results to the screen. This can be sped up significantly if specific nodes can be doing more specialized work without having to worry about post-processing the results. The Xserve is being used as a benchmark, and we hope to see the same level of performance with networked minis. The demo units need to be returned prior to the National Qualifying Event, so unfortunately we may be stuck purchasing several more minis on our own.

Tomorrow we are heading back up to the test field, where we are going to live on gas station sandwiches and bottled water for the following 72 hours leading into the site visit. The topographical map we are generating is looking very good at this point, and the replacement compass and servo speed controllers ought to noticeably improve vehicle control. I will be able to blog periodically from Appleton, so cross your collective fingers for good news!

-Dave

Pickup pick up

The pickup I bought on the cheap is turning out to be less and less of a deal. Allow me to vent:

• First of all, neither the drivers' side door nor the tailgate lock. Fortunately no one knows what the heck that big six-wheeled thing in the back is or what it's worth. And they'd need another big truck to steal it.
  
  
• The "check engine light" is perpetually on, probably because the exhaust is mostly disconnected from the engine or because of the wires hanging off the transmission sensor. Fortunately the light isn't very distracting because it has a piece of electrical tape covering it up.
  
  
• 1991 Ford F-150 Lariats have dual gas tanks: front and rear. We had been using the rear 16 gallon tank, but I decided to fill up both in preparation for a long drive up north. I ended up pumping several gallons of gasoline all over the ground through a 2" hole in the front tank. Oops.
  
  
• I had the left rear tire blow out on the highway driving home from Appleton at 12:45am last Thursday night. The truck was loaded up with all of our stuff in the back, and I was on the side of the road in the middle of the night with almost no cell phone coverage. I managed to get ahold of a 24hour flatbed towing service via 411, and after a few redials I finally gave them my location. They charged a hookup fee and I paid for the 80 mile round trip to Madison. Riding in between the two drivers on the way home, I couldn't shake this mental image of the tailgate flying open on the pickup and the last 8 months of work shooting out of the back like one of those pop guns you get on the 4th of July. We pulled in to my house at about 2:30am, where the (friendly; determined) drivers took all of the next hour trying to get the angle right to unload the truck in my driveway. The trailer hitch dug a nice gouge in the concrete and I swear all the sledgehammering woke up everyone for miles around. The pictures didn't come out so well, but I'm going to post them anyway :)
  

-Dave

Field days

I could have called this post "Field Daze" but I imagine people would have stoned me for wanton homophone overuse. Now people can stone me for saying "wanton homophone overuse" instead.

Last Tuesday we were out at the site for most of the day, working on getting consistent runs back and forth over our 200m course. We have much more varied terrain to deal with, including elevation changes and quit a bit of vegetative ground cover. We were using a new TCM2-50 compass from PNI which accounts for pitch and roll changes and makes driving the new terrain possible. We've had some trouble with acceleration throwing off the compass readings, but PNI was kind enough to send us a similar compass with a different liquid in the pitch and roll sensor which is less susceptible to lateral G forces (It should arrive any day now. It HAS to.).

We also found that the opto-isolated resistors we use for speed control do not scale anywhere near linearly. This means that if we ask for 2% throttle, we get something like 46% with nothing in between. One alternative is to use the old servo mechanism that drives a physical potentiometer, but these have the opposite problem; scaling up slowly to 50% with very little control at the top end. The third option is to go back to the original Whyachi PWM controllers and do our best to isolate the digital POT from any feedback from the Sevcon controllers. The runs on Tuesday used the servos, and the next round will test the digital POTs. Then it will be time to decide once-and-for-all which setup to use for the site visit.

Wednesday and Thursday we spent at Mike's parents' place in Appleton working on vision. We incorporated a true 3D view into our test application, which allows us to visualize the internal representation of the terrain and get a much better grasp on what needs to be changed. The first terrain maps were pretty much garbage, but by the end of the weekend we had decent model output. The most important step was calibrating the cameras, so that each level of disparity between the stereo pairs mapped accurately to a distance in the real world. More vision testing will be done at my house in Madison this week, before heading back to Green Bay on Friday to prepare for the site visit Monday morning, Aug 15th.

Pictures

-Dave

At the field today

Mike and I will be living at his parents' house in Appleton for the next few days. This morning we are heading out to Appleton Electronics to look for linear opto-isolators and hopefully a few specimens of our old friend the Xicor X9315WP, then it's up to the field for a full day of testing with the latest driving logic code. We are running the vehicle on a Mac mini now, with all dashboard controls sent wirelessly over CarNet to the monitoring laptop. Only a few changes to the driving logic to fix some glaring bugs; otherwise we want to see how it handles on the new terrain. More later.

-Dave

New test site pictures

Here are a few pictures taken over the past few days. Nothing too exciting, but you can see that the new test site is a lot more difficult than the flat field we've been using. We have trails, vegetation, and slopes at a variety of angles. Not to mention our first true hot weather test at 104 degrees last weekend. Thanks to Sampson Valley Vineyard for letting us use the land!

Test site pictures

-Dave

Report Card

All sorts of new info since last time you got a real update. Once again, in no particular order:

• Found the perfect location for testing and the second site visit


• Unfortunately, it's 2.5 hours away and our transport vehicle gets roughly 13mpg


• Discovered that our compass is much more prone to inclination error than we thought


• Fortunately, we think we found a replacement that speaks RS-232 and includes tilt sensors


• Sped up vision by almost two orders of magnitude


• Using AltiVec optimizations on the PPC CPU


• Streamlined our dashboard application and improved turning ability


• No more gingerly negotiating shallow-angle turns


• Successfully ran a waypoint following test at much greater speed than we ever have before


• Estimate approx. 15mph, will easily go faster with higher voltage. Coming soon.

Current work is being focused on replacing our servo-controlled potentiometer speed sweeps with re-engineered Whyachi devices. As you recall, we had some major issues with supplying too much voltage to the IC. Mike is testing an idea to isolate the control side of the circuit from the output side of the circuit, which would allow us to go back to controlling speed with two servo outputs instead of the currently required four.

Also, I ended up buying a welder and angle grinder. The loaner we were using several months ago is no longer available, and hey, who doesn't love a guy with a welder? The only metal work left is really the battery and computer mounts, so nothing structural. Not sure if that will get done sooner or later, but when it does, you can count on some new pictures.

I have a few pictures to post of the new test site that should go up sometime later tonight.

-Dave