Chapter 19

Defense Advanced Research Projects Agency Development Center, Nevada
September 22, 2005, 0900 Pacific Standard Time

Conversations were more animated than usual, as the MAADS development team drifted into the conference room.

Charles walked to the front of the room and said, “I think yesterday was a good test. A number of people whose opinions matter, including senior people from DARPA, told me they were impressed with the performance of MAADS. We will wait for Sergeant Jackson before we start. He joined the Rangers for their debrief and after-exercise get together, and I want to hear what he has to say.”

Charles started an informal discussion of the previous day's events by asking, “TMR4 nearly got stuck in that stream. Any suggestions for solving or mitigating the 'falling in holes' problem?”

Arjay responded, “I have this idea for a short range ground mapping radar.”

“Is there a commercially available version?”

“Unfortunately, there is no commercial application for the technology, but it should work.”

“And take five years to perfect,” Richard said.

“I know bringing a new technology to operational status takes time, but in principle ground mapping radar will solve the problem.”

Charles decided this discussion wasn’t going anywhere, and it was time to switch to a more productive topic.

“Fiona, when are we getting those fuel cells from Lawrence Livermore?”

“They promised me two for evaluation last week, but they didn’t arrive. I’ll have to chase them up.”

Charles said, “One of the DARPA people, I spoke to yesterday, said we may be able to get fuel cells out of Samsung. He gave me a name to contact. I'll pass it onto you.”

“That would be good. The more I hear about the Samsung fuel cells, the more I think they are ideal for powering the aerial vehicles.”

Sergeant Jackson walked into the room in his perfectly pressed uniform, and took his usual seat. Conversations died down as attention focused equally on Charles and Jackson.

Charles addressed him. “Sergeant Jackson, we waited for you to arrive, so we could start the meeting with a report on the Ranger's debrief. I'd like you to give us the main points, and the reaction of the troops.”

Jackson stood, walked to the front of the room, and faced the group. He consulted a set of handwritten notes before speaking.

“I'd like to apologize for being late. My commanding officer needed to speak to me.

“In yesterday's exercise, your machines caused twelve casualties among the blue force. The blue force killed four of your robots. The robots were killed after the blue force casualties occurred. Single shots caused all the blue force causalities. Forty-millimeter grenades fired from M203 grenade launchers caused all the robot casualties.

“The blue force took its casualties over a period of fifteen minutes, and infiltration by all three platoons was successfully blocked. The blue force had no idea what they were up against until the Third Platoon scouts climbed a ridge here.”

He turned to the map he had used two days earlier, and pointed to a spur of higher land that jutted into the right-hand side of the valley. “From this position the Third Platoon scouts saw one of the robots. It was moving back and forth between the same two locations. Had it been stationary, they would have probably ignored it, thinking it was just a regular ATV left by one of the Observer-Controllers, but they could see it was moving around without a driver.

“The Rangers kept the robot under observation, while a squad went to investigate. When the leading member of the squad got within range of the robot, it fired its laser weapon. That shot caused the blue force's twelfth casualty, and the squad realized it must be part of the red force. In response, the Third Platoon took it out from the ridge with simulated grenades. The other platoon leaders decided to use the same tactic, killing the other three robots they had located. At that point, the Observer-Controllers said the exercise was over.”

“What did the Rangers think of the exercise?”

“The idea of robot soldiers fascinated some of them. Others were pissed because they thought someone must have cheated to get the robots placed just right to ambush the Rangers the way they did.”

“Did anyone figure out the aerial vehicles?”

“No one. One of the men was talking about 'this bird watching them all the time', but the others just thought he was letting his imagination get the better of him. Apparently, he has a reputation for being superstitious. I mean more than the normal amount for a soldier.”

“Your reaction, Sergeant Jackson?”

“Twelve to four is a pretty good kill ratio against Army Rangers, but not exceptional for well thought out, coordinated ambushes. It’s hard to fight an enemy that only fires single shots, and gets a kill every time. Snipers are one of the most difficult opponents to beat, but against well-trained, numerically superior troops, it is usually a suicide mission for the snipers.

Richard asked, “So our 'toys' did OK?”

Jackson smiled. “Better than I expected.”

“Thank you, Sergeant Jackson. Now you have seen MAADS in operation, I would like you to give us an analysis of our weapons options. This is your area of expertise, and a subject the rest of us know little about. Let me know when you are ready?”

“I had pretty much completed my analysis before I left. I should be ready in a couple of days.”
Jackson didn't mention that he had continued working on the problem after he had left the project.

“Sammy, you are our tactical robot expert. How did they perform?”

Sammy, as usual, was well prepared. “Apart from the software problem with TMR4 that Richard should speak to, the tactical robots performed to specification. They found their way to the assigned locations, without getting stuck, rolling-over, or encountering mechanical problems, and successfully executed the ambushes. The laser rifle attachment, we connected to the targeting software, worked fine, even though the mechanism to move the rifle was just something temporary, I put together for the test.

“We rigged the robots with laser-detecting vests on the front and top, and when the simulated grenade bursts hit the four TMRs, we remotely shut them down. Everything went pretty much according to plan.”

Charles looked across at Richard who was in conversation with Jennie. “Richard, what went wrong with TMR4?”

Richard continued talking to Jennie, almost long enough to constitute a snub, before replying. “It was a dumb software bug. Are there any other kinds? It was in my code. I watched how the robot behaved, and as soon as I looked at the code, the bug was obvious. I’ve fixed it.”

Charles respected, that when Richard was wrong, he would come straight out and admit it. Never any excuses. Not that Richard was wrong that often.

Charles asked, “Sammy, can you give us your assessment of a grenade's effect on a tactical robot?”

“I can give you the engineering perspective. We haven't run any tests, but I did some research on the subject. The robots and their equipment are nothing like as vulnerable to concussion and blast as a human being is. The blast itself would be unlikely to have any effect on the all-terrain-vehicles, which are built to take shocks and impacts, and these days, electronics are robust. You can throw the ruggedized laptops we use, off the top of a five-story building, and they still work.”

Several people smiled at this statement. They had been there when the vendor had done precisely that to demonstrate his product's robustness. He claimed to have to have thrown the same laptop computer more than fifty times, and it still worked.

“Shrapnel could have an effect depending on where it hits. Before we replaced the pneumatic tires with foam filled runflat tires, they would have been the most vulnerable points, but now shrapnel will have no real effect on them.

“The vulnerable parts are the computer, networking equipment, the two cameras, and the electrical wiring and motors. A hit on the computer or electrical wiring would almost certainly disable the robot. The network adapter and aerial are also vulnerable. If they were damaged, the robot would still function, but we would be unable to communicate with it. The tactical robot would continue to perform its current mission but it couldn't receive a new or amended mission nor could we access its visual or infrared images.”

Sergeant Jackson raised his hand slightly to get Sammy's attention, and asked, “You mean grenades wouldn't have knocked out the robots?”

“Probably, not. A grenade exploding say five meters above a tactical robot might disable it maybe one time in five. That's a guess by the way. Obviously, if you get a direct hit with the grenade it's very likely to cause serious damage.”

Sergeant Jackson gave his military perspective. “Grenade rounds are not designed to directly hit the target. They are intended to burst above or close to the target, and to kill or disable by shrapnel or concussion.”

Charles said, “Thanks, Sammy. That raises an important issue, we haven't really considered, the vulnerability of our tactical robots, and the other MAADS units, to different kinds of weapons. Sergeant Jackson could you work with Sammy, and look into that for us.”

“Certainly.”

“Fiona, can you give us an update on the aerial vehicles?”

Fiona followed the example of Sammy and stayed seated. Her accent had a pleasant burr to it. Charles had asked her about when they had first met. Fiona said it was lowland Scots.

“The aerial vehicles detected the infiltrators soon after they entered the surveillance area. The C2 software directed one of the vehicles to fly a pattern covering two of the infiltrator groups, and this meant both groups were only under periodic surveillance, but the tracking software did its job, and reliably predicted the infiltrators' locations.

“After the Rangers in the center made contact with TMR3, we overrode the C2 software, and directed the aerial vehicle to fly low over the Rangers to see if it aroused suspicion on the ground. We also directed the same aerial vehicle to monitor TMR4, so we could see what was happening from the air. As a result, we temporarily lost contact with some of infiltrators on the right when they changed direction after seeing the TMR4 from the ridge. This meant they approached TMR4 without being observed by the aerial vehicle.”

Charles asked, “Did our intervention, overriding the C2 software and directing the aerial vehicle, allow the Ranger's to approach TMR4 unobserved?”

“It certainly contributed to it.”

“What would have happened if we hadn't intervened, and kept the aerial vehicle in the search pattern the C2 software directed?”

Richard interrupted to answer the question. “You intervened twice to override the C2 software, and then we encountered a bug. It's pointless to speculate what might have happened had those three things not occurred in that sequence.”

“Richard, we need to include an assessment as part of our report. Why is it so difficult to walk through the code to see what would have happened?”

Richard hesitated before answering, and Charles caught him exchanging a look with Jen. “If walking through code was sufficient to tell us how software will behave, we wouldn't need to test, and we would never have bugs. Software behaves predictably at the level of individual instructions, but to do useful things requires large numbers of instructions in combination, and those combinations often behave unpredictably.”

“So you found a bug in the code?” Charles stated rather than asked.

Charles saw Jen give a slight nod of her head. It wasn't often he caught Richard out, but he clearly had here. So much for Richard coming clean when there were problems.

“Just tell me what the software should have done, if it had performed to specification.”

Richard replied, “If the C2 software knew there was a problem with TMR4, it would have moved TMR5 to intercept the infiltrators, and that would have prevented them approaching TMR4.”

Again, Richard hesitated as if he was debating whether to say more. “We didn't find a bug. What we found, I would call an embedded assumption. If a tactical robot is not reporting something is wrong, and thinks it is proceeding toward its objective, then the C2 software doesn't detect a problem.

“TMR4 was happily moving between a rock wall it couldn't cross and a stream its software told it not to cross. As far as the tactical robot was concerned, everything was fine, and it would have continued until its batteries ran out. The C2 software needs to be able to detect when a robot is repeating behavior and failing to reach its objective, and then redirect the robot.

“You have to understand this much more than a bug. A new layer of sophistication is needed in the C2 software. It needs to react to things that don't happen, as well as things that do.”

Charles said, “Richard, just tell me, in your report, what you think would have happened if we hadn't intervened, and the software had performed to specification. And don't add any new layers to the software, without my explicit agreement.”

Charles thought that Richard needed to learn not to have these confrontations in public.

Sammy threw a softball question to Fiona to relieve the tension. “The tree cover wasn't a problem?”

Fiona replied, “It sometimes made observation patchy, but this is not dense canopy forest, and in the infrared, the aerial vehicle could see the men almost all the time.”

“Thanks, Fiona.”

Charles brought the group back to the agenda. “Richard, you have already covered the problems with TMR4, can you give us an update on the other software systems.”

“Apart from the problem with TMR4, all the software systems worked to specification, and we didn’t encounter any significant problems or bugs. The C2 software worked perfectly. We transmitted the mission data, defining the area to defend, and the software assigned surveillance areas to each of the aerial vehicles and blocking locations to the tactical robots. Once the aerial vehicles detected targets, and projected their routes, the C2 software reassessed the blocking locations, and repositioned two of the tactical robots.”

Richard looked at Fiona and then at Charles. “It's difficult to properly evaluate the performance of the aerial vehicle's tracking software because of the number of interventions Fiona made.

“I can say, we still have more work to do in determining the areas the aerial vehicles patrol, and the search patterns they use, but we always knew this would be a process of incremental improvement.”

Charles knew Richard resented people interfering, as he saw it, with the operation of his software, and as usual had trouble letting go of an issue.

Charles normally avoided taking sides or criticizing people in public, but he needed to move Richard past the issue.

“Fiona did the right thing by making those interventions. They were things we needed to evaluate. You will just have to factor them out in your report.”

Charles paused, looked directly at Richard to make sure he got the message, and said, “It's not only robots that are prevented from reaching their objective by unproductively repeating behavior. Now, can we move on to the performance of the tactical robot's software.”

After a momentary hesitation as Charles' point hit home, Richard said, “The tactical robot's search and targeting software worked to specification. Everyone should watch the images as they search for, and then kill the targets. It makes fascinating viewing. A shot to the center of the torso caused every infiltrator casualty. The accuracy of the targeting is impressive, even at close to maximum range, and when the targets were crawling on the ground through undergrowth.

“I have put together a compilation of images showing how the tactical robots find targets, aim their weapons, and then kill their targets.”

He typed a command on his laptop, and a moving visual image of the forest displayed on the screen behind him. The image periodically stopped, and went to high magnification showing a patch of forest undergrowth in detail.

“The robot is performing its normal search for targets. There are none in view, so each time it finds an indication a target may exist and zooms in for a closer look, it comes up negative.

“The search software processes the data coming in from the cameras in order to identify people. It uses multiple strategies to find anything in an image that might be a person or indicate the presence of a person. When it finds one of these indicators, both cameras zoom in to get a closer look.

The search software then monitors the places where it finds indicators, in order to get enough data for a definite identification. When the search software positively identifies a person, it notifies the targeting software, which then takes over. You have to keep in mind, this is an ongoing process with more data being added all the time.

The projected image continued its moving, long distance view of the forest, and periodic magnification for a closer look. A magnified visual image lingered on a patch of forest undergrowth.

“I’ve read through the log, and can summarize what's happening. The search software has found strong indications of a human head. The head has now disappeared from view, but scrutiny of the spot and the immediate area will stay high on the software’s priority list. The software has directed the infrared camera to zero in on this area, but because the head is no longer visible it can’t find any infrared indicators of a human. By the way, I have edited out the ongoing search of other areas.”

Richard stopped talking while visual and IR images alternated.

“Although both cameras scan together, they operate independently. In order to see what's happening simultaneously with both images, I've windowed one image within the other. When the visual image displays, you can see the infrared image in a window at the bottom right. When the main image switches to the infrared, the visual image shows in the window.”

The visual image suddenly went to high magnification, followed almost immediately by the IR image in the window.

“The search software has found a strongly positive set of indicators. It's very likely a human is there.

“I’ve taken the images you have just seen, and zoomed them up to the maximum useful magnification. The visual image is not that clear because of the man's camouflage outfit and paint, but the IR image is very clear. Even a human observer would recognize this as the head and upper torso of a man, and I assure the software is much better than a human observer.”

Charles thought, ‘what Richard is saying is, his software performed much better than a person would because he wrote it.’ The man was insufferable, and quite possibly right.

Richard continued with his commentary. “The search software uses visual and infrared images. Both are effective for different reasons and work well in combination. Visual image searching uses a more sophisticated strategy because it's been around longer, and the software is more developed, but infrared search has better potential.

Once the search software has positively identified a human, the targeting software takes over. It works exclusively using infrared, and takes over control of the infrared camera.”

A high-magnification infrared image appeared on the projection behind Richard. Superimposed on the image were rifle crosshairs.

“This image is from the targeting software. The crosshairs are a literary device I added because the mechanism the software uses is not something I can show on a screen.”

A recognizable partial view of man’s head and torso appeared in the infrared image.

“Normally, the software targets either the head or the torso, with the head preferred. But because we used a laser rifle, and the targets were wearing laser-detecting vests, I had to put in some temporary code to just target the torso.”

A red flash momentarily blocked out the infrared image. When it returned, the man was rolling away from the camera.

Richard said, “The red flash was another literary device. It shows the point the targeting software fires the laser rifle at the man you can see in the image. He rolls away when his vest emits a sound indicating a hit.”

Sergeant Jackson found Richard's self-congratulatory tone, irritating, and wanted to get some useful information. He asked, “You referred to 'the maximum range'. What is the maximum range of the equipment?

“The range of the equipment is dependent on the search software and the targeting software. The targeting software only operates once the search software has positively identified a human target, so in practice the range of the targeting software is limited to the range of the search software.

“Because a human target, in most circumstances, is so distinctive on the infrared, the targeting software works, out to the limit of the IR optics. With our equipment that's about two hundred meters, and we could extend it further with better optics but there's no point because the range of the search software is much less for several interrelated reasons.

“The search software processes all the images it receives, looking for indicators. As we increase the magnification we create more images to process, and consequently, the search software makes fewer checks on each image.

"For the purposes of this discussion, the total number of checks the software can perform in a time period is fixed, and determined by the power of the computer processing the images. So at greater distances searching is less thorough, and only progressively more obvious targets are found. Better optics wouldn't make much of a difference, but faster computers would.

“Up to eighty meters, I would rate the search software as highly effective, and up to a hundred and twenty meters, as moderately effective.

“I hope, I don't have to explain to the people in this room that there is an inverse square relationship between distance and surface area.”

Sergeant Jackson had no idea what an 'inverse square relationship' was, but the range of the equipment was useful information. “So the maximum effective range is between eighty, and one hundred twenty meters?”

“For the tactical robots as they currently exist, yes.”

Charles asked, “Richard, I know you are still working on target classification, but can you describe your current thinking?”

“We explored several solutions to the 'shooting only the bad guys problem', starting with the straightforward. Give the good guys a device that identifies them as a friend so that MAADS knows not to target them. Something that is inevitable in my view as automated weapon systems are introduced to land warfare. These devices are called 'identification, friend or foe’, IFF for short. IFF solves only part of the problem. What happens if there are civilians in a combat zone?

“The search software can recognize armed men. At least when they are armed with a rifle or pistol. Other weapons, like grenades, are more difficult to recognize.

“I did some work on getting the software to differentiate uniformed and non-uniformed people, and also people in different uniforms, with reasonable success, but that doesn't help when the enemy is irregulars out of uniform. I originally thought uniform recognition was a useful way to help identify targets, but I now realize it's to easily spoofed and have abandoned that avenue.

“I came up with a decision tree to classify whether a person is a legitimate target or not that operates in the context of variables we set remotely. The most important variable. I've called the threat level. The threat level attempts to encapsulate the military risks and determines how aggressively MAADS behaves.

“The threat level has three settings, low, medium and high.

"Low is intended for when MAADS is in not in a combat zone, and it is a passive observer sending back images of what it sees. When the threat level is set to low, MAADS will not take any hostile actions.

"If the threat level is set to 'medium' then MAADS targets armed men not wearing an IFF. It will ignore an unarmed man unless he approached one of the units. If he did, MAADS would fire a single warning shot. If he continued to approach, MAADS would kill him.

"A high threat level is intended for combat zones, where there are unknown threats and any person armed or unarmed without an IFF can be considered hostile.

"Normal operation would use low or medium threat level. The high threat level would only be used in special circumstances such as when weapons, MAADS doesn't recognize are used. "

Charles asked, “How accurate is MAADS in determining if a man is armed? It's clearly a crucial issue.”

Richard said, “I took two hundred hours video of armed and unarmed people, civilians and soldiers, and used these videos both to train the software and to measure how accurate the software is in recognizing armed men. One of the nice things about software systems is you can keep improving them or getting them to improve themselves.

I've now got the software to the point false positives are almost zero, but false negatives are significant, at least 10 percent.”

Sergeant Jackson asked, “What does that mean?”

“It means that when the software identifies a person as a target, it is almost always right, but when it identifies a person as not being a target, at least one time in ten its wrong.”

Sergeant Jackson said, “Real soldiers have the same problem.”

Charles decided to ignore the emphasis put on the word 'real'. Are there anymore more questions?”

Arjay asked, “Charles, why don't the aerial vehicles transmit the location of infiltrators directly to the tactical robots, so the robots know where they are before the software searches for them?
Surely, it would help the robots find them faster.”

“A good question and one I asked Richard a few weeks ago. Richard, can you tell everyone the reasons you gave me?”

“There are several answers to that question, and I'll restrict myself to what I consider the two most important. Having the aerial vehicles directly transmit target locations to the tactical robots and then having the robot's search software use that data, couples the performance of the robots to the performance of the aerial vehicles. If the aerial vehicle crashes or we redirect the vehicle somewhere else for some reason, we impair the robots' ability to find targets. We could get around that problem by having two modes, one where the robot was receiving target data, and the other where it wasn't, but that adds significantly to the software's complexity and hence development time.

“The second answer, is that you have to ask yourself what benefit accrues from transmitting target data from aerial vehicles to tactical robots. The answer is, the robot finds targets more quickly, which in practice means it can target and kill them when they are further away. So, the question becomes how much sooner and how much further away, and how much military value does that have. I coded up a rough model and it said, on average a target would be killed about two seconds sooner and ninety-one meters away, as opposed to an average of eighty-nine meters away.

“Perhaps, Sergeant Jackson could tell us the military value of killing an enemy two seconds sooner and two meters further away at a range of ninety meters?”

Jackson controlled his annoyance at being asked an apparently simple, but in reality unanswerable question. “Two seconds is unlikely to make any difference at that distance.”

Charles continued speaking before Richard could take his explanation further. “That’s enough for this morning. This afternoon, starting at two p.m., I'd like talk about new capabilities. Also, I'd like to do a tour of the labs on Friday afternoon and everyone should be prepared to give a quick run down on their current activities, then hopefully people can get away early for the weekend.”

Chapter 20