Brad Ideas

Robocars & More

For the first time, a car company has put a date on shipment of a car with self-driving ability.

According to British site Auto Express, Mercedes has revealed that their 2013 S-class will feature self-driving. Not clear if there is an official company press release, though the company has been talking about such features, as have many other companies. Realize that the 2013 model year is just a year away.

I shoot with the Canon 5d Mark II. While officially not a pro camera, the reality is that a large fraction of professional photographers use this camera rather than the Eos-1D cameras which are faster but much bulkier and in some ways even inferior to the 5D. But it's been out a long time now, and everybody is wondering when its successor will come and what features it will have.

Each increment in the DSLR world has been quite dramatic over the last decade. There's always been a big increase in resolution with the new generation, but now at 22 megapixels there's less call for that. While there are lenses that deliver more than 22 megapixels sharply, they are usually quite expensive, and while nobody would turn down 50mp for free, there just wouldn't be nearly as much benefit from it than the last doubling. Here's a look at features that might come, or at least be wished for.

Better Pixels

More pixels may not be important, but everybody wants better pixels.

• Low noise / higher ISO: The 5D2 astounded us with ISO 3200 shots that aren't very noisy. Unlike megapixels, there is almost no limit to how high we would like ISO to go at low noise levels. Let's hope we see 12,500 or more at low noise, plus even 50,000 noisy. Due to physics, smaller pixels have higher noise, so this is another reason not to increase the megapixel count.
• 3 colour: The value of full 3-colour samples at every pixel has been overstated in the past. The reason is that Bayer interpolation is actually quite good, and almost every photographer would rather have 18 million bayer pixels over 6 million full RGB pixels. It's not even a contest. As we start maxing out our megapixels to match our lenses, this is one way to get more out of a picture. But if it means smaller pixels, it causes noise. The Foveon approach which stacked the 3 pixels would be OK here -- finally. But I don't expect this to be very likely.
• Higher dynamic range: How about 16 bits per pixel, or even 24? HDR photography is cool but difficult. But nobody doesn't want more range, if only for the ability to change exposure decisions after the fact and bring out those shadows or highlights. Automatic HDR in the camera would be nice but it's no substitute for try high-range pixels.

Video & Audio

Due to the high quality video in the 5D2, many professional videographers now use it. Last week Canon announced new high-end video cameras aimed at that market, so they may not focus on improvements in this area. If they do, people might like to see things like 60 frame video, ability to focus while shooting, higher ISO, and 4K video.

Last week, new studies came back on the California High Speed Rail project. They have raised the estimated cost to $99 billion, and dropped the ridership estimate to 36.8 million and $5.5 billion in annual revenue. Note that only around 20 million people currently fly the SF to LA corridor -- they expect to not just capture most of those but large numbers of central valley trips.

Congestion on the roads has a variety of sources. These include accidents of course, reductions in road capacity, irrational human driving behaviours and others, but most of all you get congestion when more cars are trying to use a road than it has capacity for.

That's why the two main success stories in congestion today are metering lights and downtown congestion charging. Metering lights limit how fast cars can enter the highway, so that you don't overload it and traffic flows smoothly. By waiting a bit at the metering light you get a fast ride once on the highway. Sometimes though, especially when the other factors like accidents come into play, things still gum up.

Now that more and more cars are connected (by virtue just of the smartphone the driver carries if nothing else) the potential will open up for something else in congestion -- finding ways to encourage drivers to leave a congested road.

A little self-plug. I have an article on an introduction to panoramic photographic technique the November issue of Photo Technique with a few panos in it. This is old world journalism, folks -- you have to read it on paper at least for now.

I'm just back from the "ITS World Congress" an annual meeting of people working on "Intelligent Transportation Systems" which means all sorts of applications of computers and networking to transportation, particularly cars. A whole bunch of stuff gets covered there, including traffic monitoring and management, toll collection, transit operations etc. but what's of interest to robocar enthusiasts is what goes into cars and streets. People started networking cars with systems like OnStar, now known in the generic sense as "telematics" but things have grown since then.

The big effort involves putting digital radios into cars. The radio system, known by names like 802.11p, WAVE and DSRC involves an 802.11 derived protocol in a new dedicated band at 5.9ghz. The goal is a protocol suitable for safety applications, with super-fast connections and reliable data. Once the radios in the car, the car will be able to use it to talk to other cars (known as V2V) or to infrastructure facilities such as traffic lights (known as V2I.) The initial planned figured that the V2I services would give you internet in your car, but the reality is that 4G cellular networks have taken over that part of the value chain.

Coming up with value for V2V is a tricky proposition. Since you can only talk to cars very close to you, it's not a reliable way to talk with any particular car. Relaying through the wide area network is best for that unless you need lots of bandwidth or really low latency. There's not much that needs lots of bandwidth, but safety applications do demand both low latency and a robust system that doesn't depend on infrastructure.

The current approach to safety applications is to have equipped cars transmit status information. Formerly called a "here I am" this is a broadcast of location, direction, speed and signals like brake lights, turn signals etc. If somebody else's car is transmitting that, your car can detect their presence, even if you can't see them. This lets your car detect and warn about things like:

• The car 2 or 3 in front of you, hidden by the truck in front of you, that has hit the brakes or stalled
• People in your blind spot, or who are coming up on you really fast when your're about to change lanes
• Hidden cars coming up when you want to turn left, or want to pass on a rural highway
• Cars about to run red lights or blow stop signs at an intersection you're about to go through
• Privacy is a big issue. The boxes change their ID every minute so you can't track a car over a long distance unless you can follow it over every segment, but is that enough? They say a law is needed so the police don't use the speed broadcast to ticket you, but will it stay that way?

It turns out that intersection collisions are a large fraction of crashes, so there's a big win there, if you can do it. The problem is one of critical mass. Installed in just a few cars, such a system is extremely unlikely to provide aid. For things like blindspot detection, existing systems that use cameras or radars are far better because they see all cars, not just those with radios. Even with 10% penetration, there's only a 1% chance any given collision could be prevented with the system, though it's a 10% chance for the people who seek out the system. (Sadly, those who seek out fancy safety systems are probably less likely to be the ones blowing through red lights, and indeed another feature of the system -- getting data from traffic lights -- already can do a lot to stop an equipped car from going through a red light by mistake.)

They say that famous deaths come in threes. That's no doubt just an artifact of our strange sense of coincidence, but after Jobs and Ritchie, tonight we learn of the death of John McCarthy, AI pioneer and creator of LISP.

Since getting involved with Google's self-driving-car team, I've had to keep silent about its internals, but for those who are interested in the project, a recent presentation at the intelligent robotics conference in San Francisco is now up on youtube. The talk is by Sebastian Thrun (overall project leader) and Chris Urmson, lead developer. Sebastian led the Stanley and Junior teams in the Darpa Grand Challenge and Chris led CMU teams, including BOSS which won the urban challenge.

The list of robocar teams grows again with a new project from Oxford university, led by Paul Newman. Nissan is also involved, though the base vehicle is a Bowler Wildcat off-road vehicle.

Earlier I wrote about the transportation potential of walking robots-of-burden like BigDog. While these robots are not for the long haul, a whole range of options are opened up by a wheeled vehicle that can get to where the road ends, and then lower legs to walk along rough terrain, up stairs and up hills.

Boston Dynamics has gone even further with their latest model, AlphaDog

I'm actually not a fan of login and sessions on the web, and in fact prefer a more stateless concept I call authenticated actions to the more common systems of login and "identity."

But I'm not going to win the day soon on that, and I face many web sites that think I should have a login session, and that session should in fact terminate if I don't click on the browser often enough. This frequently has really annoying results -- you can be working on a complex form or other activity, then switch off briefly to other web sites or email to come back and find that "your session has expired" and you have to start from scratch.

There are times when there is an underlying reason for this. For example, when booking things like tickets, the site needs to "hold" your pending reservation until you complete it, but if you're not going to complete it, they need to return that ticket or seat to the pool for somebody else to buy. But many times sessions expire without that reason. Commonly the idea is that for security, they don't want to leave you logged on in a way that might allow somebody to come to your computer after you leave it and take over your session to do bad stuff. That is a worthwhile concept, particularly for people who will do sessions at public terminals, but it's frustrating when it happens on the computer in your house when you're alone.

Many sites also overdo it. While airlines need to cancel your pending seat requests after a while, there is no reason for them to forget everything and make you start from scratch. That's just bad web design. Other sites are happy to let you stay "logged on" for a year.

To help, it would be nice if the browser had a way of communicating things it knows about your session with the computer to trusted web sites. The browser knows if you have just switched to other windows, or even to other applications where you are using your mouse and keyboard. Fancier tools have even gone so far as to use your webcam and microphone to figure if you are still at your desk or have left the computer. And you know whether your computer is in a public space, semi-public space or entirely private space. If a browser, or browser plug-in, has a standardized way to let a site query session status, or be informed of session changes and per-machine policy, sites could be smarter about logging you out. That doesn't mean your bank still should not be paranoid if you are logged in to a session where you can spend your money, but they can be more informed about it.

Some time ago I ranted how I hate being invited to fill out a customer service survey after every bloody transaction in the online world. Companies don't seem to clue in that if they ask for surveys too much, they end up getting the opinions only of the angered or overeager.

Some Robocar updates, since with Burning Man and Singularity U my posting volume has been down:

Yesterday, Don Henley (of the Eagles) penned an editorial in USA Today supporting the Protect IP Act (PIPA) which has serious free speech implications and turns web sites into copyright police. Don called out both the EFF (of which I am a Director) and Google (which is a consulting client of mine) so I have this whimsical response for him:

Today an op-ed by John Sununu and Harold Ford Jr. of "Broadband For America" (a group of cable companies and other ISPs which says it is really a grass-roots organization) declared that the net needs a better pricing model for what Netflix is doing. For a group of ISPs, they really seem to not understand how the internet works and how pricing works, so I felt it was worthwhile to describe how things work with a remarkably close analogy. (I have no association with Netflix, I am not even a customer, but I do stream video on the net.)

You can liken the internet to a package delivery service that works somewhat differently from traditional ones like the postal service or FedEx. The internet's pricing model is "I pay for my line to the middle, and you pay for your line to the middle and we don't account for the costs of individual traffic."

In the package model, imagine a big shipping depot. Shippers send packages to this depot, and it's the recipient's job to get the package from the depot to their house. The shippers pay for their end, you pay for your end, and both share the cost of creating the depot.

Because most people don't want to go directly to the depot to get their packages, a few "last mile" delivery companies have sprung up. For a monthly fee, they will deliver anything that shows up at the depot addressed to you directly to your house. They advertise in fact, that for the flat fee, they will deliver as many packages as show up, subject to a fairly high maximum rate per unit of time (called bandwidth in the internet world.) They promote and compete on this unlimited service.

To be efficient, the delivery companies don't run a private truck from the depot to your house all the time. Instead, they load up a truck with all the packages for your neighbourhood, and it does one delivery run. Some days you have a lot of packages and your neighbours have few. Other days you have few and they have a lot. The truck is sized to handle the high end of the total load for all the neighbours. However, it can't handle it if a large number of the neighbours all want to use a large fraction of their total load on the same day, they just didn't buy enough trucks for that, even though they advertised they were selling that.

This is not unreasonable. A majority of the businesses in the world that sell flat rate service work this way, not just internet companies. Though there are a few extra twists in this case:

• The last mile companies have a government granted franchise. Only a couple can get permission to operate. (In reality -- only a few companies have got permission to have wires strung on poles or under the street.)
• Some of the last mile companies also used to be your exclusive source for some goods (in this case phone service and TV) and are concerned that now there are competitors delivering those things to the customers.

The problem arises because new services like Netflix suddenly have created a lot more demand to ship packages. More than the last mile companies counted on. They're seeing the truck fill up and need to run more trucks. But they proudly advertised unlimited deliveries from the depot to their customers. So now, in the op-ed, they're asking that companies like Netflix, in addition to paying the cost of shipping to the depot, pay some of the cost for delivery from the depot to the customer. If they did this, companies would pass this cost on to the customer, even though the customer already paid for that last mile delivery.

As I prepare for Burning Man 2011, I realized I had not put my gallery of regular sized photos up on the web.

Much earlier I announced my gallery of giant panoramas of 2010 which features my largest photos in a new pan-and-zoom fullscreen viewer, I had neglected to put up the regular sized photos.

Tuesday we and Aneesh Chopra, CTO of the USA come to Singularity University and among many things, he was asked about immigration. (In part because our class comes from 35 countries and many of them would love to be entrepreneurs in the USA.) Chopra announced some immigration rule clarifications that had come out that day which will help things somewhat. They did rule clarifications because getting congress to do meaningful reform is very hard.

A report from China Daily details a successful 286km trip by a vehicle from the National University of Defense Technology.

This is a challenge to blog readers to come up with (or find examples in practice) of good systems to allocate students to parallel sessions based on their preferences. I've just concluded one round of this, and the bidding system I built worked OK, but is not perfect.

The problem: Around 80 students. On 10 days over 4 weeks they will be split into 3-5 different parallel sessions on those days. Many sessions have a cap on the number of students, and more students will have them as a 1st choice than can fit. Some sessions can take many students and won't fill up. The students can express their preference as ranking, or with numeric values.

This is known in the literature as the Allocation problem, and there are various approaches, though none I found seemed to fit just right, either being easy to code or having existing running code. But I am keen on pointers.

The goals:

• Maximize student satisfaction/minimize disappointment. Giving a student their 1st choice is good. Giving 3rd or 4th choices is bad.
• The system must be easy for the students to understand and use.
• Fairness. This has many meanings, but ideally mismatches that can't be avoided should be distributed. If somebody gets a 4th choice one day, they perhaps should have a better shot at a 1st choice on another day.
• It's nice if there's a means of applying penalties to students who violate rules, sneak into sessions etc. Academic violations can result in less chance at getting your 1st choice.
• It should be flexible. Sessions may have to be changed or many not fully finalize until a week before the session.
• It is nice to handle quirks, like duplicated sessions a student takes only once, but where the student might have preferences for one of the instances over another. There may also be pre-requisites, so only students who take one session can have the sequel.
• Things change and manual tweaking can be advised.

Rank sessions in order, 1st come, 1st served

This was used in the prior year. Much like a traditional sign-up sheet in some ways, students could indicate their choices in order. If more students had a session as 1st choice than would fit, the ones who filled out their form first got in. This gave priority over all 10 days and so it was changed to rotate each week to distribute who was first in line.

A wrapup of robocar news from the past couple of weeks:

Nevada governor Brian Sandoval rides in Google Car

After Nevada's recent legislation directing their DOT to explore legal operations for robocars in the state, the governor "took the wheel" of a Google car. Very positive impressions from the governor and DMV head.