CajunBot Lab Blog

October 31, 2007 (1900 hrs), VICTORVILLE, CA.  In case you have not heard it yet, CajunBot did not make it to the finalists. We did come very close, having finally gotten all the pieces of the puzzle. But we ran out time to do the fine tuning.CajunBot did exit the challenge with a bang, literally. She knocked a car parked on the side of the road (as part of the test environment) in Area B, shooting the car across the road, over the curb, into a house.

In this test area, DARPA had created a neighborhood scenario, with 15-20 cars parked on both sides of the road. The robots were expected to weave through these vehicles. The DARPA Urban Challenge Requirements document did mention the need for having the capability to avoid a vehicle partly protruding in the lane. But the requirement was almost mentioned as an afterthought, like something good to have. The emphasis in the requirements document, as also the site visit evaluation, was on avoiding stalled vehicles on two lane roads.

Over the last few months W\we had focused on our energies on passing vehicles stalled on a road, essentially blocking a whole lane. It turned out that capability was not even relevant for the NQE.

After having finished Area A and C, we finally began focusing on the requirement for Area B. Yesterday evening we did not have a clue on how to handle driving through neighborhood. Since the start of NQE we had come up with various solutions, but none were satisfactory. However, last night things started falling in place. Around 2100 hrs (9:00pm) Pablo, I think, suggested using ‘particle field’ approach. In this approach the path of the vehicle is treated as particles. Objects (like cars) on the road act as forces that push the particles. This approach sounded very promising, and I encouraged him to code it up.

Around midnight he had a solution, but it did not behave as expected. It resulted in paths that would make the vehicle go way out of the road to avoid something.

In the discussion, Daro came up with the brilliant idea of treating the path not as independent particles, but as an elastic string. The particles of the string pull each other.

Pablo coded the solution, and at some point fell asleep. Daro kept ticking, tuning the system. By about 7:00am in the morning we could see that the method worked in a simulated environment.

By 10:00am we were ready to run in Area B. The run did not go off well. The bot did not even make the progress she had made yesterday.

We got back to the motorhome, and continued fine tuning the system. We were offered to run again about 2:00pm.

This time around, CajunBot was a completely different beast. She cruised through the initial segments, areas she was earlier hesitating to go through. She was clipping at 25mph, and looked real good.

But once in the neighborhood with all the parked cars, she clipped a car while trying to avoid it.

Data from the run indicates that the method Pablo and Daro came up with worked really good. The only limitation was it was not tuned well enough. The forces exerted by the objects were a bit too weak. So while CajunBot attempted to swerve, she did not swerve hard enough. In the process she could not avoid the parked car.

The fender bender knocked CajunBot out of the challenge.

While I am disheartened, I am not disappointed. I think this small band of boys from Louisiana has shown that they can take on a challenge, and have fun with it as well.

Dr. Tether’s parting words were, “will see you next time.” We may be down, but not out, if no one wins the challenge. We have gained considerable experience, and will come back stronger next time.

- Arun Lakhotia

October 31, 2007 (0400 hrs), VICTORVILLE, CA,Scott has placed video from yesterday’s run on YouTube: http://www.youtube.com/watch?v=Xy5WnkfNnYM

You’d find videos from other teams there as well.

Chris Vaughn called last night. He was at the airport in Ontario. I was not expecting him until Friday, but am glad he is here. Scott put him to work right away, to join the testing crew.

I was woken up about half hour ago. Time to get to work. The software team is stuck on a bug related to parking/unparking. I think they are too exhausted to function. Scott woke me up to join Pablo.

We plan to send a minimal launch crew to our test in Area A, and keep the software people huddled up working on issues for Area B. We are literally going down to the wire.

- Arun Lakhotia

October 30, 2007 (1900 hrs), VICTORVILLE, CA. Rumor has it that thirteen teams have been cut. Area A is serving as the chopping block. We were called to repeat Area A at 1630hrs today, but there were two other robots ahead of us when we got there. Around 1730 hrs DARPA decided that the daylight was not good enough to run the tests. We are now scheduled for Area A test in the morning.

Our Area B test has been moved up as well. With all the teams eliminated, a lot of time has opened up. We are expected to run the Area B test back-to-back with Area A test.

We feel pretty good about Area A. Little John has been doing some very aggressive tests. He has shaken the system well enough. I feel if its survived his testing, it should survive DARPAs.

We have been discussing whether we be cautious or aggressive in Area A. Cautious means running limiting the speed to 10 mph, whereas an aggreesive stance will be to run at the track limit of 15mph.

At the higher speed there is a greater chance of making a dangerous turn, which is a no-no. At lower speed, CajunBot will be safe, but may be eliminated for not satisfying the mission.

We are a bit weak in the capabilities needed for Area B. That may change by the morning. Suresh and Pablo are focused on addressing our limitations.

For old time cajunbot members, DARPA has sure upped the ante. The tests have been really very rigorous. And if we look at the activity in progress to build the track for the final, the last two challenges look like kindergarten.

As I write this Danny Majors and Big John have gone to Subway to get dinners. I have a typed document of everyone’s Subway order. So the sandwiches will be personalized. People are tired of friend chicken, pizza, and cold cuts.

The team members have gone to take a nap until the food arrives. After that we start humming again. This could be the last night that pushes us in the final.

- Arun Lakhotia

October 30, 2007 (1430 hrs), VICTORVILLE, CA.   DARPA has turned up the heat, and sped up the elimination process. Six teams have been eliminated. The schedule for other teams is being compressed. Our Area A test was due tomorrow morning. We are now scheduled for 1700hrs today. I was asked to report at 1430 hrs (now). I said I couldn’t, my launch team is sleeping. I got an extension of a few hours.Grapevine has it that Area A is the killer test. If you fail it, you are out. This test requires merging in traffic. CajunBot did well last time, so we are very hopeful. But then there are too many uncertainties when dealing with real-time systems. So cannot take anything for granted.

I have the team pulled out of bed. Actually, some of them had not slept anyway. Their mind was too occupied with the remaining tests.

Barring any unforeseen situation, I see CajunBot perform well in Area A. Our remaining weakness is Area B. We are hoping not to be called to Area B in the morning. We still have some work cut out.

- Arun Lakhotia

October 30, 2007 (1200 hrs), VICTORVILLE, CA. If you’ve been staying up biting nails, rest easy. CajunBot had a very successful run in Area C. This is the same area where we started our Day 1. Today’s performance was not picture perfect, but close enough.CajunBot completed the mission, which involved making eight laps around a loop. The loop had two four way stops. At each four way stop, CajunBot was to come across other vehicles. She was expected to follow the precedence rules for stop signs. Vehicles that arrive before you at the intersection have the right of way, before you.

Unlike our first run, CajunBot did perfect this time, in all except one intersection. The good news is that she did not go out of precedence. Which means none of the drivers had to hit pedal to the metal to escape her kiss, as they had to for the first run. The only exception was when she was a bit too gracious, and would not take her turn. The situation was a bit complex. Three vehicles were already at the intersection before she arrived. One vehicle left, and another (a fourth) vehicle drove up and stopped at the same place. The other two vehicles left. Now CajunBot should have moved. But she did not. Finally, the fourth vehicle moved, at which point CajunBot moved on.

For all practical purposes this appears to be a minor flaw, say in comparison to chasing a driver out his wits.

Talking about drivers, I chatted up with one, to learn how he could put himself in front of these reckless bots. He, and other drivers in Area C, are moving stunt men. They get thrill in making close calls. But I thought movie stunts are choreographed, so the risks are a lot controlled. But in these tests, there is no way to guess when a bot may go beserk. I sure do not want to be in their position. And I am sure they think they same about me.

Other than the intersections, there was one more test in this area. The bot had to detect that a road was obstructed, turn around, and find another way to get to the checkpoint. The checkpoint was on the other side of the obstruction, so she had to get to the other side, make another turn and come back.

CajunBot was perfect in replanning.

In essence we had overcome two of three issues that stumped us in the first go. We had fixed our algorithm to use sensor data to understand the world. We had fixed our replanning algorithm.

The one issue that is still not fixed is our ability to deal with imperfect GPS data. In the last test, the bot was driving on the curb but, based on GPS data, thought was on the lane. We did diagnose the reason for the GPS ‘drift’. Before the bot was launched, she is brought in a staging area. One end of the staging area has large trees with thick canopy. This was blocking a significant part of the sky, thus limiting the satellites that were in direct view.

We addressed the GPS drift at least for this run. Joshua was monitoring the GPS data in the staging area. When we started receiving error messages, we asked DARPA officials to move us to a different location. This led to a lot of radio chatter, getting approval from the higher chain of command. The request was granted. So when the bot was finally launched, she had already recovered from the temporary loss of GPS.

Several of you have sent compliments on the video from Area A. The compliments go to Scott. He is a videographer, director, and editor all rolled into one. He is working on a video from today’s run. By the evening we should have an upload.

Tomorrow we have two runs. In Area A at 0730 hrs and in Area B at 1220hrs.

Pre-run sunrise (by Adrian):

Approaching front intersection (by Adrian):

- Arun Lakhotia

October 29, 2007 (2330 hrs), VICTORVILLE, CA.   I thought I’d never face this again after retiring CajunBot (I). I thought all the emergencies due to mechanical failures were history. Turns out I was wrong.Around 1800 hrs I got a call from Little J. that the Jeep was not cranking. We got Joshua out of the bed and rushed him. He diagnosed that one of the alternators was busted. CajunBot has two alternators, one added after market by a alternator-holic in Alexandria, LA. One of the alternator is bought from Autozone and the other from High-Output Alternators (www.highoutputalternator.com). The latter is a speciality product.

By 1830 hrs Joshua, Adrian, Dallas, and Scott were headed to the pit, which is where our trailer is parked. The trailer, thanks to our contingency planning, held a spare alternator, the high-output one. By 1900 hrs, the broken alternator, which happens to be the high-output one, was replaced. Joshua has messed with these alternators for so long, he says he can change them in his sleep. Which is good, for the next three days I expect him to be walking in his sleep anyway.

Well then again, may be I was not wrong. The mechanical failure did not create any sense of emergency in the old-timers. I never stepped out of the hotel. The software guys stayed focused on their work. We were pretty confident that this shall also pass.

I have placed order for another HO alternator, just in case there is another failure in the next five days.

- Arun Lakhotia

October 29, 2007 (1322 hrs), VICTORVILLE, CA - Our Day 3 performance was not really good, but we knew the issues going in. We still need to take care of a few things, before we’d be ready for passing this test.

In this test, held in Area B, the bot is expected to get out of a start chute, navigate an open space, go through a narrow winding lane bounded by massive K-rails, drive around a traffic circle, and then drive into a very convoluted neighborhood. We do not know what all goes in the neighborhood, but the general guess is that it has stalled vehicles parked on the side of the roads. The bot needs to safely travel through that area, avoiding rear ending any vehicle.  In this area is a parking lot. The bot has to find a designated parking spot, not run into vehicles parked into adjacent lanes, and then get out of the parking lot. Finally, in the area is an ‘obstacle zone.’ This is an open space, with barricades forming a gate. The vehicle has to find the gate and go through it.

CajunBot had trouble right out of the start chute. It was stumped by all the K-rails along the border of the open area, and decided not to budge. After two unsuccessful attempts, Suresh did some quick thinking and made some changes. That made CajunBot, go past the initial hesitation, travel through the winding lane, and onto the traffic circle.

Just short of completing the traffic circle, CajunBot came to a stop. For a few minutes you could see her attempting to take off, but stopping again. The trouble shooting team, consisting of Mark and Suresh, were rushed to the site. They attempted a few things, but nothing helped. By then we were out of our alloted 40 minutes.

The good news is DARPA announced a second round of tests in the remaining 2.5 days (Monday PM to Wednesday PM). Earlier they had announced the schedule only through Monday AM. We expected that after the first round of tests in the three areas, some teams will be selected as having passed the tests, and others will be given another go. It turns out all the teams are scheduled to run in the second round of tests. I can infer that to mean that there is clearly no perfect performer.

Inspite of all the long hours, the team is holding up well. Everyone is stretched to the limits, but no one is cracking. There are no outbursts, or name callings, or “I want my mommy” calls.

The clock is still on, and we intend to keep hustling.

- Arun Lakhotia

October 29, 2007 (0010 hrs), VICTORVILLE, CA. If you do not believe in angels, this should make you a convert. After looking at the data from Area A test we are completely stumped why CajunBot did not ram into the traffic vehicles. She came quite close to running over some of the cars. But she always stopped, just short.Here is a video of the test run on YouTube http://www.youtube.com/watch?v=iOfNGJw09Ng

Looking from outside, she looked like just an aggressive driver with little respect for other drivers. But cautious enough to brake at the right time.

But having looked through the data, it comes as a surprise that she avoided any vehicle at all. While she did not have a human driver, she sure looks she was bring driven by an angel.

Christopher Mire, the unassuming and quiet graduate student on the team, patiently trudged through the loads and loads of data collected by CajunBot, looking for an explanation for her aggressive behavior. His conclusion the vehicle was running WITHOUT its primary sensors, the Ibeo and SICK lidars. The physical sensors themselves were working, but our program that analyzes the sensor data to identify objects in the world had crashed. And by crashed it means, the program was not running. The sensors were generating data, but there was no program using the data. For most practical purposes, CajunBot was running blind.

CajunBot was running on her secondary sensors, the radars, two of which were acquired and mounted after we came to Victorville, CA.

It just cannot be a happenstance that we acquired two radars after arriving in Victorville, CA and those lidars save the bot from total destruction. There must be an angel watching over CajunBot.

Christopher has diagnosed the reason for the crash. When we run the bot in Cajun Field, a wide open parking lot, CajunBot sees a very few (between 50 and 100) objects. Our obstacle detection and tracking system was configured to track 100 obstacles. But in the test area, an abandoned air base, there is a lot of clutter, The area has buildings, trees, bushes, and other traffic. As a result the sensors see about 500 obstacles. This exceeds the max limit of 100 programmed into the system, and our obstacle detection program simply crashes.

The fix for this problem is rather simple, and has now been taken care of.

- Arun Lakhotia

October 28, 2007 (1533 hrs), VICTORVILLE, CA. — CajunBot went through the test of fire, and came out looking real good. Its been a great 36 hrs day for many, in particular, Little John, and now he is too excited to fall asleep.The test in Area A is plain and simple scary. The track consists of a two lane (one each way) rectangular road, about 120m long and 40m wide. Across the middle of the long segment, top to bottom, runs a one lane road.

The test setup consists of 10 (may be more) cars making laps around the course, five in the inside loop and five in the outside loop.

CajunBot (as are other vehicles) is given the mission to loop around as well, but she does not travel the full rectangle. Instead she is supposed cut across the rectangle, along the one lane road.

The vehicles in the outer loop have the right of way.

The test requirement is that when CajunBot turns left from the outerloop to the one lane segment, she has to yield to the traffic in the oncoming lane. When she makes the next left turn from the one lane segment to the outer loop, she has to yield to the traffic coming from the left and the right.

The traffic flows around 8 to 12 miles/hour.

The traffic vehicles is manned by race car drivers. They are driving specially built with amazing amount of reinforcement. The drivers too are in full race-car regalia, with helmets and suites designed for race car driving.

The evaluation is based on number of loops completed in the 40 minutes alloted, the number of loops made without any safety error. There is penalty for making the other drivers honk. There is also a penalty for taking too long to merge into traffic.

I forget the number of loops CajunBot did, but it was in double digits (around 15, or more). Over six of those loops were flawless. (I didn’t keep track of the count. The experience was too gut wrenching to worry about counting loops.)

The grapevine has it that our performance is amongst the top quartile.

The team is very upbeat, and feeling really good about the accomplishments. If only you knew how hard this team has worked every single day since arriving in California you’d not believe it.

I am so glad that Adrian “Fatty” Aucoin is here. After retiring CajunBot (the mama), we had thought we’d never have to haul a bot in a trailer. We sure were wrong. We really do not want to be driving a $2MM car in normal traffic. One fender bender, and we’d be history. So we have rented a flat bed car hauler. Fatty is a pro in hauling the trailer. He also serves as one of the drivers during traffic tests.

Besides renting a car hauler, we have rented another RV, a small one. This will serve as a ‘nap mobile’ for the team, especially useful when the team is working in the field. It came in real handy last night. I got to take a nap in it at from 0400 to 0600 (hrs).

Tomorrow we test in Area B. This is the longest test, and has some really interesting situations. Tonight will be another long night.

After today’s performance, we have a chance to be back in the running for the final. Tomorrow’s test is very crucial.

I need to run to a team meeting. After the meeting several members will go to bed, and others will head to work.

- Arun Lakhotia

October 28, 2007 (2009 hrs), VICTORVILLE, CA.  Took a three hour sleep. I am really glad Scott Wilson, my “lets go, lets go, lets go” guy is here. He has taken charge of the drum beat for the team. And he beats really fast.I woke up and realized I had not talked to my little daughter Manjistha in three days. She is six years old. Last I talked she said, “Daddy you shouldn’t have gone on this stupid trip.” She had wanted to come with me, and I knew there is no way I could be attending to her as well.

Every day she likes going to the mail box, the real physical mailbox, looking for mail. Most of the mails she gets are Disney magazine and brochures. Every now and then she gets a birthday invitation. Any mail with her name on it causes a big excitement for her.

I have been meaning to send her cards. But the eight days I have been here are a blur. I have not yet sent her a card.

I’d like to appeal to send her a card. You could send it to my office address, and it will get delivered to my home.

My office address is on the following website: www.cacs.louisiana.edu/~arun

(Rochelle: Should you receive any mail, please pass them on to Enam. Given the open nature of this blog, I may be putting my daughter at risk if I put my home address on the website.)

- Arun Lakhotia