Squat orange robots and a set of adaptive algorithms are making it possible to deliver online orders faster. The system, so far installed in two giant Staples warehouses, allows workers to fill two to three times as many orders as they could with conventional methods. The startup that developed the robots and software, Kiva Systems, based in Woburn, MA, announced yesterday that it is rolling out a third system, for the pharmacy giant Walgreens.

Kiva Systems' CEO and founder, Mick Mountz, likens the system to random access memory chips. The warehouse is arranged in a memory-chip-like grid composed of rows and columns of freestanding shelves. The grid gives robots access to any product in the warehouse at any time. The robots serve two basic functions. First, they deliver empty warehouse shelving units to workers who stock them. The workers might stock one unit with a mix of paper, various types of pens, and computer software, all compiled from large pallets that had been delivered to the warehouse. Then, when a consumer submits an order, robots deliver the relevant shelving units to workers who pack the requested items in a box and ship them off. "We turn the whole building into a random access, dynamic storage and retrieval system," Mountz says.

If a consumer orders an item at 2 P.M. on a Thursday, he says, at 2:01, a robot can be delivering that order to a worker to pack. If an order has multiple items, robots will line up for workers as fast as the workers can pack the items. Once the items are packed, robots can pick up the boxes, storing them temporarily or delivering them to the appropriate delivery truck.

Mountz says that the system allows workers to fill orders much faster than conventional systems do because the robots can work in parallel, allowing dozens of workers to fill dozens of orders simultaneously. In one type of conventional system, workers have to walk from shelf to shelf to fill orders, and all that walking takes time. With the Kiva system, several robots can be dispatched to collect all the items in an order at once. The robotic system is also more efficient than conveyor-based systems, in which elaborate conveyors and chutes send boxes past workers who pack them as they go by. In such a system, the slowest part of the line, which could be the slowest worker, limits the overall speed. With the help of the robots, each worker fills an entire order, so one worker doesn't slow everyone else down.

The robotic system is also faster because the entire warehouse can adapt, in real time, to changes in demand. Robots move shelves with popular items closer to the workers, where the shelves can be quickly retrieved. Items that aren't selling are gradually moved farther away. More-conventional warehouses can also be adaptive, Mountz says, but it takes much longer to rearrange items.

Shopping online from the comfort of your desk chair is certainly easier than traveling to a store and lugging home heavy bags. But for all its effortlessness, online shopping falls short when it comes to finding something you weren't looking for but would like to buy. Recommendation systems, such as those built by Amazon, try to uncover these gems, but many fall short of appropriately catering to an individual shopper.

Now a Seattle-based startup called Cleverset thinks it has the secret to the next-generation recommendation system: a type of computer modeling found mainly in artificial-intelligence research labs. Cleverset's system weighs the importance of the relationship among individual shoppers, their behavior on the site, the behavior of similar shoppers, and external factors such as seasons, holidays, and events like the Super Bowl. Using these ever-changing relationships, Cleverset's system serves up products that are statistically likely to match what the customer will find interesting.

Online retailers can have millions of products in their warehouses, but a consumer only has a limited view of what's available when she comes to a site, says Bruce D'Ambrosio, Cleverset's founder and a professor of electrical engineering and computer science at Oregon State University. "You've got gigabytes of stuff behind your website," he says, "and you only have a megapixel of display." The challenge for most online stores is finding the best products and information to show in that tiny space on the screen.

Recommendation systems have been around for nearly as long as online retail sites have existed, and each varies slightly in its approach. Many systems just match products to people by looking at the products that others have bought. For instance, if you are looking at a blender, and people who bought the blender also bought a toaster oven, then the system would suggest a toaster oven to you. The problem here, says D'Ambrosio, is that all this analysis of purchases happens offline, and the system has no awareness of what a consumer is trying to accomplish at that specific point in time.

Cleverset uses an approach called statistical relational modeling, developed in the past decade, in which each piece of information in a data set is linked together based on its relationship to every other piece of information. This contrasts with the previous view of looking at data as if in an Excel spreadsheet, where everything carries an equal weight.

Statistical relational modeling has, for the most part, stayed cooped up in research labs. It's been used to develop technologies such as natural-language processing (to extract relationships from text), bioinformatics (to find relationships between genes and proteins), and computer vision (to help robots see scenes as collections of related items). Daphne Koller, a professor of computer science at Stanford University, says that statistical relational modeling is good in these instances because there is a lot of uncertainly within the data sets. Relationships can be established, she says, and then statistics must be used to determine the likelihood and importance of each relationship.



In the case of Cleverset, the system starts collecting data and forming relationships within that data the instant a person hits the retailer's website. D'Ambrosio says that, as with many site analytics tools, Cleverset relies on little programs that retailers install on their websites. These programs can track the previous site that the consumer viewed, and if it was a search engine, it logs the keywords used. As the user clicks on items, Cleverset's system creates a more detailed view of his interests and compares it with those of other people using the site. What sets the system apart is that it organizes customers' behaviors into a data set that includes information on how various behaviors relate to each other. The system also pulls in outside information, such as whether or not a person is shopping during a Super Bowl commercial break.

While Cleverset was founded in 2000, its technology has only recently reached the point at which the results are good enough to make a significant difference in the competitive e-commerce industry. D'Ambrosio says that sites that use Cleverset--which include Overstock.com and Wine Enthusiast--experience, on average, a 20 percent increase in revenue per customer. The company is also earning some media buzz: when Cleverset presented its technology at the Web 2.0 Summit in San Francisco last month, it came away with two audience-voted awards: "Best in Show" and "Most Likely to Exit First."

Stanford's Koller says that a recommendation system such as Cleverset's "fit neatly into the framework of statistical relational modeling because it's all about relationships." She argues, though, that it might be impossible to make a single system fit every kind of e-commerce site. For instance, Netflix, which launched a competition to build a better system, uses different methods than a site that recommends clothes. (See "The $1 Million Netflix Challenge.")

Cleverset works with each site to tailor its technology appropriately, says D'Ambrosio, which will be important, as the company soon plans to launch with a number of undisclosed "very large retailers" that bring in $100 million or more annually. D'Ambrosio adds that the technology is still improving, and he and his team see future versions of their system including even more input from merchandisers about how their customers use their site.

Everyscape, a startup based in Waltham, MA, is getting in on the rush to create a virtual version of the real world. Although the site will launch this fall under the shadow of mapping giant Google Earth, Everyscape's cofounders say that users will find the company's look and feel quite different. "We're working on a human experience," CEO Jim Schoonmaker says. "Google has built a superhuman experience."

Everyscape's demo opens in the middle of San Francisco's Union Square, below the Dewey Monument. Users can choose the auto-drive mode, which gives a virtual tour of the area's sights and shops, or they can explore on their own. Auto drive orients a user by showing her the general layout of Union Square before taking her into Harry Dentin's Starlight Lounge and bringing her out again for a dizzy, swirling look at the night sky above the Dewey Monument.

The site is designed to give a full immersive experience. A user should be able to tour Union Square virtually, Schoonmaker says, and then feel comfortable navigating it in real life.

Google Earth, in contrast, opens with a satellite's view of the earth resting in space. From there, users can fly down to explore chosen terrain or look out at the stars. While many areas are created with flat satellite photos, some locations include links to street-level photos taken by users. Images showing a 3-D view of certain buildings can also be layered onto the map, using a special programming language called KML.

In Everyscape, building interiors are constructed the same way as the rest of the environment: by stitching together a series of panoramic photographs taken by company photographers or contributed by users. Within each photograph, a user can swivel through a full sphere of motion. To move users from within one panoramic photograph to the next, Everyscape's servers estimate the locations of the cameras in each photograph and use that information to build sparse 3-D geometry that forms the building blocks for an animated 3-D transition. Everyscape CTO and founder Mok Oh says that the transition works because it simulates people's real-life attitude toward moving from place to place. "Getting there is not what you want," he says. "Being there is what you want."

Derek Hoiem, a researcher at the University of Illinois at Urbana-Champaign who designed the technology behind the 3-D site Fotowoosh, says that 3-D immersive sites are popular now because of their appeal to users। "When you're able to control the environment, it feels more lifelike," he notes. While Hoiem says that Everyscape's technology gives a good approximation of motion, he also says that he would like to see greater freedom of movement, rather than just swiveling and transitioning.

Ironically, the original version of Everyscape's technology, used by the first company that Oh founded, Mok3, had the type of capability that's on Hoiem's wish list. Mok3 built software that can use panoramic photographs to generate environments interactive enough for a game engine, and that looks much like a walk-through captured on video. Infinite Corridor animation.) In search of a business model, Oh scaled back the technology to make it transfer more easily over the Internet and founded SuperTour Travel, which created interactive environments to show off high-end hotels and other travel destinations to potential customers. With Everyscape, Oh hopes to use what he learned with SuperTour to virtually reproduce the entire world.

In a business model based in part on SuperTour's, Everyscape plans to make money by helping businesses build their interiors for a fee. Schoonmaker says that he expects shopkeepers to understand the need to virtually display their physical inventory and store layout. "That's where all their money went," he says. "That's what they need to show you." In the future, Schoonmaker hopes to add more interactive features to help businesses function virtually. Future additions might give users the ability to buy merchandise inside a store with the click of a mouse, or might add a virtual maître d' that could help visitors make dinner reservations at a restaurant and recommend items on the menu.

Everyscape plans to launch this fall with environments for parts of San Francisco, Boston, and New York. Other future plans include adding user-controlled avatars and features for mobile devices.

Computer-generated effects are becoming increasingly more realistic on the big screen, but these animations generally take hours to render. Now, Adobe Systems, the company famous for tools like Photoshop and Acrobat Reader, is developing software that could bring the power of a Hollywood animation studio to the average computer and let users render high-quality graphics in real time. Such software could be useful for displaying ever-more-realistic computer games on PCs and for allowing the average computer user to design complex and lifelike animations.

Adobe is focusing its efforts on ray tracing, a rendering technique that considers the behavior of light as it bounces off objects. Since it takes so long to render, ray tracing is typically used for precomputed effects that are added to films, computer games, and even still pictures before they reach the consumer, explains Gavin Miller, senior principal scientist at Adobe.

With the rise of multicore computing, Miller says, more consumers have machines with the capability to compute ray-tracing algorithms. The challenge now, he says, is to find the best way to divvy up the graphics processes within general microprocessors. "Adobe's research goal is to discover the algorithms that enhance ray-tracing performance and make it accessible to consumers in near real-time form," Miller says.

Consumer computers and video-game consoles compute graphics using an approach called rasterization, explains John Hart, a professor of computer science at the University of Illinois at Urbana-Champaign. Rasterization renders a scene by generating only those pixels that will be visible to a viewer. This process is fast, but it doesn't allow for much realism, explains Hart. "Rasterization is limited in the kinds of visual effects it can produce, and has to be extensively customized just to be able to approximate the appearance of complicated reflective and translucent objects that ray tracing handles nicely." For instance, in real life, if a light is shining at the side of a car, some of that illumination could reflect off metal in the undercarriage, and this would create a reflection on the ground that's visible to a viewer who's looking at the car from above. Rasterization would ignore the pixels that make up the undercarriage, however, and the reflection would be lost.

Ray tracing takes a fundamentally different approach from rasterization, explains Miller। "Rather than converting each object into its pixel representation, it takes all of the geometry in the scene and stores it in a highly specialized database," he says. This database is designed around performing the following fundamental query: given a ray of light, what points on a surface does it collide with first? By following a ray of light as it bounces around an entire scene, designers can capture subtle lighting cues, such as the bending of light through water or glass, or the multiple reflections and shadows cast by shiny three-dimensional objects such as an engine or a car.

Essentially, then, ray tracing tries to find the right information in a database as quickly as possible. This isn't a problem for rasterization, says Miller. Usually, the rendering process is straightforward, and data is cached and ready to go when the processor needs to use it. With ray tracing, however, the brightness of any given point on a surface could have been created from multiple bounces of a light ray, and data about each bounce of light tends to be stored in a separate location in the database. "This is a nightmare scenario for the caching strategy built into microprocessors, since each read to memory is in an entirely different location," says Miller.

He explains that his team is exploring various approaches to making these database queries more efficient. Previous research has produced algorithms that bundle certain types of data together to simplify the querying process. For instance, bundles of data can include information that represents rays of light that start from roughly the same location, or rays that head in nearly the same direction. Adobe is not releasing the details of its approach, although Miller says that his team is trying to find the most efficient combination of database-management approaches. Once the researchers develop software that can effectively manage the memory of multicore computers, then ray-tracing algorithms can be rendered at full speed, he says.

"Adobe makes software that improves a user's ability to create and communicate visually," says Hart of the University of Illinois. "Software like Photoshop provides methods for processing photographs, but by adding ray tracking, users will have the ability to create photorealistic images of things they didn't actually photograph." One of the biggest obstacles at this point, he says, is making the system work fast enough so that a user can run a ray-tracing program interactively.

The current ray-tracing approach alone won't solve all the problems that computer-graphics researchers are tackling, Hart adds. It's still impossible to perfectly simulate the human face. "This is an elusive goal," he says, "because as we get more realistic ... subtle errors become more noticeable and, in fact, more creepy. Once we get faces right, we will need high-quality methods like ray tracing to render them, and we'll want it in real time."

The system is still just a research project, and the company doesn't provide a timeline for when it might make it to consumers, but technology on all fronts, including advances in multicore architecture, is advancing rapidly. Miller suspects that consumers will start to see real-time ray tracing in products within the next five years.