Building Stanley

This knol is about Stanley, the robotic car that was used by the Stanford Racing Team which under the leadership of Sebastian Thrun and Mike Montemerlo won the Darpa Grand Challenge on October 8th, 2005. I will focus all my attention on the car itself, for an in-depth view of the complete system design see Stanford's excellent white paper[5].

The car itself was built at the Electronic Research Lab (ERL) of Volkswagen of America (Palo Alto, CA) and I had the great pleasure to lead the team of engineers[6] who brought it to life.

The background story

It all started when one of my colleagues, Martin Lienkamp, paid us a visit at the ERL. On the plane ride from Germany he had read an article about the 2004 Grand Challenge and was convinced that "we (Volkswagen) can do better than that" referring to the fact that none of the autonomous vehicles that had entered that race were able to complete the course. His comment stuck with me and I thought this would be an exciting challenge. The ERL was a small team at that time, and we did not have the AI and autonomous driving expertise that was available in other parts of our company. An obvious solution to this shortcoming was to partner up with one of the many great US research groups, and I decided to call on a friend for advice on this matter. Joseph O'Sullivan was working at Google at that time and had in a past life roamed the hallways of the robotics program at Carnegie Mellon University. After I explained my idea to Joseph, he mentioned that probably the best person to work on this just happened to be local. Sebastian Thrun had been Joseph's PhD adviser at CMU and had recently moved to Stanford. A few emails later I was meeting  Mike and Sebastian at Stanford's Artificial Intelligence Lab. The timing was near perfect, as they had just started thinking about putting together a team to compete in the next event. It was relatively easy to convince ourselves that we were a good match: Stanford would build the brains, and Volkswagen would provide the body.

All that was left to do on my side was to set the corporate wheels of approval in motion, and on June 29th 2004 I received a note that confirmed a nod by Dr. Bernd Pischetsrieder, then CEO of Volkswagen. The next day Sebastian, Mike and I committed to our collaboration. Stanley was thus "conceived" on June 30th 2004, 15 months and a few days before race day.

What's in a name?

Why Stanley? Stanford was originally planning to use a Ford SUV for this project, and Mike thought it would be clever to call the robot "Stan" (Stanford, get it? I thought you would). Obviously this joke did not work with the VW Touareg, and the team chose "roadrunner" as the new codename. In fact most of Mike's original code still refers to "roadrunner". A few months later, when the time came to have our first press event we revisited the issue of naming. Both Pamela Mahoney (our liaison with major sponsor MDV)  and myself felt that "Roadrunner" did not sound likeable enough, and there was also a brief consideration of trademark issues (Plymouth produced a Road Runner). Reminiscing the early Stan I proposed Stanley, which was deemed to sound more friendly, "human" and somehow recalled the pioneering/exploratory dimension of this project. And so the car was christened. Anecdotically, in the summer of 2005 Joe and I built a duplicate as back up, that we dubbed "Stanlette". It was a near perfect copy, with a few minor differences in the skidplate and bumper guard, an improved electric system and a more modern engine (the 6-cyl TDI). For the record, we never had to use our backup and it is the original Stanley, the same one that survived thousands of miles of testing that conquered the Darpa Grand Challenge.

Selection of the vehicle

The model of car to be used was an easy choice. The Touareg had just been released by VW, and its offroad capabilities were unmatched by any "luxury SUV". The engine was chosen according to two criteria: efficiency and reliability. Efficiency was important because it:

1. spared us the complexity and failure modes of an additional fuel tank (the race could last up to 10hrs)
2. saved money and clean air over the hundreds of hours of testing by our engineers

I probably don't need to delve much into the reason why reliability is paramount, but if you followed the race closely, you probably know that one of our strongest competitor was critically flawed by engine troubles. If you need to ask, they were not using a Volkswagen.

The engine chosen that fulfilled these criteria was a 5-cylinder diesel powerplant that had been proven in many VW models (for gearheads, the engine is known as "R5"). The R5's power output was similar to Hummer H1 engine, using only a fraction of the fuel. It has a large alternator that could sustain the electrical power needs of the robot (stack of computers, sensors, mechanical actuators, etc.)

Making Stanley 

Left to right: Sven Strohband, Joe van Niekerk, Cedric Dupont and Charles Markey.

Like many designs, it all came together in bits and pieces. Torque specs were gathered from 4AM discussions (yay timezones) with Joerg Kreft, a researcher in Wolfsburg, Germany, and the steering linkage design was inspired from various chats I had with our in-lab mechanic Joe Bajurin. The team we put together for this project was made of relatively new additions to the ERL, Joe van Niekerk - the master of controllers who could build anything out of a custom PCB and a microcontroller (just left of me in this picture);  Philippe Allessandrini - who would turn my steering system sketches into an actual manufacturable design; Eric Jensen - who programmed the maxon controller; and Charles Markey (far right on this picture) - who put this all together in record time, delivering a usable Stanley to Stanford within weeks of the car's arrival on US soil.

The nose sticking in at the far left of the picture belongs to Sven Strohband, my colleague who brilliantly took the team to the starting blocks after I left for business school a few weeks before the race.

 

Steering 

The main steering actuator was a chain drive directly attached to the steering column. We chose a chain drive because it had the twin benefit of flexibility (yes, even a Touareg flexes a little in off-road conditions) and the option of completely removing the linkage (using a master link in the chain), therefore offering "hardware security" when disengaged, allowing us to drive manually on public roads with confidence.I will insert a picture of the initial sketch if I find it, but the final design consisted of a welded steel bracket holding an electric motor (maxon with planetary gear reduction). The motor's shaft was attached to a motorbike sprocket, driving a chain connected to a second sprocket that was fastened to the steering column. In order to attach the second sprocket to the steering column without dismantling the car completely, we split it in two (yes these are made of hardened steel, Charles spent a few blades on this one...) and fastened it to the steering column with custom CNC'd fasteners that would engage in the existing grooves. This design proved extremely robust and reliable and did not experience a single failure.

Details of the steering actuator, the motor and primary chain drive are visible in the foreground. The encoder cable is on the bottom left.

 

Of course a good design is nothing without brilliant execution, and Charles Markey should get all the credit for the surgical adjustment of the contraption we had put together.

Acceleration

Like many modern cars, the Touareg's gas pedal is basically a potentiometer, measuring the deflection angle applied by the driver's foot and turning it into an electrical signal transmitted to the Engine Control Unit. This made it relatively easy for us to control acceleration by computer by faking the behavior of a regular gas pedal.

Braking

The Toureg's brake booster is equipped with an electrically controlled valve that we controlled directly to provide breaking force. The great benefit of the design is that this breaking mechanism is completely parallel to actually depressing the brake pedal, so a human pilot could save the day at any time by slamming the brake if our silicon brain skipped a beat and the car were to veer off the trails. This was somewhat tricky because of the overly smart ABS controller which would sometimes notice our abuse and force the vehicle into "safe mode", a mode that reacts to severe self-diagnosed problems and is designed to allow you to reach a mechanic at very limited speed so your car can get fixed. It took Lars's hacking genius to keep this from happening unpredictably.

Below is a very early diagram that shows the initial design of the full drive-by-wire system (note that it is missing the gear selector). If you have really good eyes you'll notice the basic principle of splicing the signal wires from original controllers (ABS module, gas pedal) and applying our alternate controls through a simple switch panel. This allowed for easy manual override by a test pilot.

 

Gear shifting

Stanley was able to shift gears using a simple linear actuator, and a few custom CNC'd fasteners. Although it was not used during the race  (it was considered a non-critical failure point) you can see a cool video of it in action here. 

Stanley growing up

It was quite a ride from the time I first picked up Stanley at the Home Depot parking lot in East Palo Alto until it turned into the autonomous champion that took away the prize on October 9th 2005 (technically the victor was awarded on October 10th, the morning after the race). Below are a few pictures illustrating a few important stages the robot went through.

The first picture shows Stanley pretty much as "shipped" (actually, it flew) from Germany. This is me testing its off-road capabilities in Arizona. The only non-stock part is a set of high performance off-road tires. If you look closely you can see one of the GPS antennas sticking out in the back (the UFO shaped thing on the roof).

First trip to the desert under the supervision of Mike. Notice the GPS antenna (center rear and front), 4 SICK laser scanners, and a basic camera mount. Stanley came back with no scratches.

Stanley with its armor (the skid plates are installed, and you can tell the holes in the front bodywork where the bumper guards will fit). The roof structure is a simple contraption made of extruded aluminum profiles that provided flexible mounting options for the various sensor systems. It also fitted the required siren and orange revolving light. The brilliant "Drivers not required" tag line was thought up by my colleagues Jackie Lai and Joyce Chen, along with most of the other design elements.

Early control setup, notice the classic analog joystick controlling the 2-ton beast. Nifty! On the second view, you can see the steering mechanism protruding a couple inches. 

 

Stanley's final brain. The power module is on top of the intel blades (center), the actuator controllers are on the left column, two additional yellow top batteries (used as buffer) are clearly visible in the spare tire bay. The round thing at the bottom left is an air reservoir for the Touareg's adjustable air suspension:

 

Stanley passing its first exam: Darpa's on-site inspection. Yes, it is driverless! I was driving an interceptor vehicle, just in case...

A new paint job to match Stanley's big (or at least fast) brothers from the Paris Dakar. You grew up so fast... (and you look smoking hot). Notice the "Sunfuel" decal on the fuel hatch: Stanley was a zero fossil-carbon emission vehicle during the race as its fuel was synthesized from "rapsoel". Of course the special diesel had to be shipped in from Germany so it was not zero-impact, but we proved the concept.

Warming up the engine... sharing my last few moments with Stanley before he takes off for the race

Measuring up the competition. From right to left: H1lander, Stanley, Sandstorm

A few moments away from taking home the $2M prize (picture by Darpa official during the race)

Other stuff I found digging around in my old notebooks

First notes and designs of the steering mechanism:

 

 

 

Early sketch of the drive-by-wire system I made for a fax to a German colleague:

The Stanford Racing team at the unofficial testing headquarters, the /X ranch in the Mojave Desert:

How Stanley sees his environment (random screen capture of a test run - you can see more examples in the white paper[5]):

Hanging out with the Axion Twins... (long story):

Lars, fillin 'er up with Sunfuel - a synthetic diesel made from plant oil:

A (very) happy bunch: Lars, David Stavens, Mike Montemerlo. Stanley came back a little dusty, but without a scratch:

Acknowledgements

A few colleagues at Volkswagen I would like to explicitly acknowledge:

Sven Strohband: When I went off to business school two months before the race, Sven jumped in and took over my duties with brio, bringing it home for the team. This was a huge task, and Sven did everything right to bring Stanley to the starting blocks in great shape.

Carlo Rummel: My boss, who was crazy enough to let me go ahead with this, and did not flinch when the conservatives where nagging about the riskiness of our endeavor to VW's reputation. Without his courage to ignore these voices and continuous front line support for the project we would have failed.

Lars-Erik Jendrossek: The Touareg Guru. As noted before, when it comes down to it, it is the car that won the race, and Stanley's reliability was largely due to Lars' care. He cured Stanley of all the wounds our multiple electronic and mechanical hacks had inflicted upon it.