None of the usual academic players have indicated interest. I've checked with the robotics departments at MIT, CMU, and Stanford. Other competition may appear, but a big field is unlikely.
Rules and constraints
The rules of the competition are straightforward, but worth close reading.
350 miles is a very long offroad trip. The problem will be finishing, not going fast.
There will be a maximum time, and we don't know what it is yet. So it's hard to work out a minimum speed. Achieving enough speed offroad to complete the course before dark will be very tough. We should assume for now that night driving will be necessary.
A first cut at an approach
This is my (Nagle's) first cut at the problem.
The basic problems are those of low-level AI: not falling over and not bumping into stuff. Above this is directing the vehicle towards a goal.
The rules say that the course can be traversed by a standard truck. So I propose using a commercial pick-up type truck or related vehicle, with some modifications for better offroad performance. We need a forgiving off-road vehicle. Jeep-type vehicles roll over too easily if driven badly offroad. A Hummer 1 (or HUMMV) is very stable, but the $100K+ cost seems excessive. A Hummer 2 ($50K) might be an option to consider.
GMC trucks from 2003 onward have an onboard stability control system called StabilTrak. This uses rate gyros and accelerometers to improve vehicle stability by independently braking all four wheels, and modulating the throttle in some situations. Other manufacturers have similar systems, but GMC seems to be the only one putting them in trucks so far. We'd like to tap into that system and get some additional outputs from it, so it can tell our systems when to slow down or steer less aggressively.
Input from people with offroading experience is needed here.
Physically controlling the vehicle is likely to be a huge hassle in terms of mechanical fabrication. DoD is developing something called the Common Robotic Kit, which is basically a system for making various DoD vehicles, including the HUMMV, into R/C cars. The bidder conference is next week, after which we should be able to find out who the vendors are. We might be able to buy a solution to this problem. More on this later.
Control system architecture
We only have a year, and limited resources. We have to use as much off the shelf technology as possible. There's a fair amount we can draw upon. Most of the major problems have been more or less solved. But nobody has ever driven such a distance autonomously without a human standing by to intervene now and then.
The most relevant paper on the subject is Emergent Architectures: A Case Study for Outdoor Mobile Robots which is Jay Gowdy's doctoral thesis at CMU. Gowdy took the off-road and on-road driving systems from different NavLab projects and made them play together. His insights on that system integration problem are definitely worth reading. Some key points:
Recognizing is hard, but depth from stereo yields to brute-force compute power. We finally have the MIPS to do this in software at adequate frame rates. More on this later. For now, see Point Grey Research and Tyzx. Point Grey does stereo vision in software; Tyzx does it in hardware. Tyzx isn't ready yet; the custom ICs are only sampling. Point Grey has been shipping for years, with the systems getting faster as CPU power increases. I was ready to order a Point Grey system today, for about $2200 for three cameras, but they can't ship - their supplier of imager chips (Sony) is behind, having favored camcorder manufacturers over them during the holiday season, they claim.
Depth from stereo is useful at short ranges. At longer ranges, we may need depth from motion stereo.
Depth from stereo is only a first step. We'll need a near-area world model, probably of the Moravec certainty grid type as used in the CMU Navlab offroad work, to decide where it's safe to drive.
We may be able to get some of the Navlab code from CMU, to use as a base for our own work.
Auxiliary safety systems
The Eaton VORAD truck anti-collision system offers a promising way to significantly improve the survivability of the vehicle. This is a proven product used on thousands of heavy trucks right now. It's a millimeter radar system with multiple radars, a control unit, and an interface to brakes, transmission, and engine that slows a vehicle down when it's about to hit something. Data outputs are also available. Price is in the $2000 range.
This is a definite win as a backup for the video-based system. The resolution of the radar is too low to be used to sense terrain. It's not useful for steering, although some side-obstacle sensing is available. Think of it as a sort of "super-bumper".
Useful question: can it see a barbed-wire fence? Being a radar, it probably can, while the vision system probably can't until it's quite close. Having two completely different classes of sensor has advantages.
Worth looking into is whether sensors from that system could be used to prevent running off cliffs or into ditches.
Related question: do we need to be able to back up? If so, we need sensors on the back, although probably only ones suited for very slow speed operation. Commercial ultrasonic back-up obstacle sensors might be sufficient, coupled with software able to retrace, in reverse, a short path recently driven forward. If we can back up two vehicle lengths, that's probably good enough. Faced with an obstacle discovered only at close range, we can then back up and turn to bypass it.
Inertial system, like Crossbow, for reliable tilt and acceleration sensing.
We will need a magnetic compass and may need a heading gyro. More on that when we look at the navigational system.
Temperature sensing on everything. All vehicle gauges visible to control system.
Good-quality onboard video recording, for debugging and publicity.
We need to be driving around by mid-2003. Maybe not very well, but driving.