摘要：机器人萝丝和詹姆斯正在做煎饼。这是一项艰难的工作，因为它需要协调力、规划以及能够很好的使用锅铲。过程中，萝丝慢慢的挪动锅铲并成功将煎饼翻转过来，人群中爆出阵阵欢呼声。
然而萝丝和詹姆斯似乎不能很好的完成这项煎饼任务。机器人应该是灵活的，但是由于那些煞费苦心精心制作的软件说明，让它们的行动犹豫不决。如果给萝丝和詹姆斯一份不同的食谱，那他们表现也会受到影响。
这是机器人一个老问题，可能只有一个解决方案，过去，机器人专家们经常让机器人执行一些高度专业化的任务，像煎饼。但是现在他们转向人群希望能让机器人掌握一些普通的技能。通过实验让人们操控机器人或者在互联网上操作虚拟的机器人，他们希望创造出灵活快速的机器人。“众包”是切实可行的，它能让机器人为人类做更多有用的事情。
TWO robots named Rosie and James are making pancakes. It's a tough task that requires coordination, planning and a fine touch with a spatula. Midway through the process, Rosie slides the spatula under the half-cooked cake, lifts the pancake from the heat and carefully flips it over. The crowd breaks into cheers.
Can I get you anything else?
Rosie and James aren't quite the accomplished cooks they appear to be, however. As far as robots go, they are as flexible as they come, but they move haltingly, and their complex routine is the result of painstakingly crafted software instructions. Give Rosie and James a different recipe, for example, and their performance would suffer.
It's an age-old problem in robotics, but one that might just have a solution. Roboticists traditionally go to great lengths to coax automatons to perform highly specialised tasks, like cooking crêpes. But now they are turning to the crowd for help in giving robots more general skills. By allowing people to pilot real or simulated robots over the internet in trial experiments, they hope to recreate the fast and flexible behaviour that comes effortlessly to humans. "Crowdsourcing is a really viable path toward getting robots to do things that are useful for people," says Chad Jenkins, a robotics researcher at Brown University in Providence, Rhode Island.
The idea is inspired by successes in other areas of artificial intelligence. For example, online translation systems are trained on pairs of documents that have been translated by humans. By comparing translations with originals, the software learns how to translate words and phrases between different languages.
A similar approach could lead to better human-robot interactions, says Sonia Chernova at Worcester Polytechnic Institute in Massachusetts. To collect information on teamwork, social interaction and communication, she and colleagues created Mars Escape, an online game in which two people each control an avatar, one human and one robot, and work together to rescue objects from a doomed Martian research lab. After Mars Escape went online last year, Chernova logged the dialogue and action from over 550 sessions of the game.
The researchers first had to throw out unusual records, like dialogue between players who traded obscenities rather than working together. Then they looked for common patterns in the data, such as methods that players frequently used to retrieve objects, and phrases they exchanged when doing so. By having software watch how people tackled the game, the software learned how to work with a human. The technique could also find a use in the games industry (see "Build a better baddie").
The real test came last January, when Chernova and colleagues mocked up a real-life version of the Martian lab at the Museum of Science in Boston. Visitors were paired with a robot powered by software based on the Mars Escape data. The results were encouraging, Chernova and colleagues say in a paper to be presented next week at the International Symposium on Robot and Human Interactive Communication in Atlanta, Georgia. Sixteen out of 18 visitors worked with the robot to complete the game and most said the robot behaved rationally and contributed to their success.
Jenkins had similar success with a proof-of-principle experiment. Last year, he wired up a wheeled robot for online access and invited people to guide it through a simple maze. Over 270 people took up the challenge. He used the data they generated to build a navigation algorithm that allowed the robot to complete a maze it had not seen before.
His next experiment is more ambitious. His lab has a state-of-the-art PR2 - the same class of robot as James - that it plans to make available online. The robot will be placed in a kitchen and users will be invited to help it perform common tasks, like fetching objects from cupboards. The data they generate could help create better domestic robots, says Jenkins. The online interface will be demonstrated to researchers this August and should be available to the public by the end of the year.
The initial experiments have also flagged up some potential problems. Players in the real-life Mars Escape complained that the robot had poor communication skills, for example. This may be because the real robot often prompted different behaviour to its virtual version. For example, some visitors issued commands to move a specific distance. No players in the online game issued similar instructions, so the robot had no appropriate data to draw upon.
If such problems can be tackled, the technique has potential, says Jenkins. Many researchers focus on domestic tasks, but people in the outside world might prioritise other uses once they get control of robots. He draws an analogy with the early days of the internet: researchers built a data-sharing system and did not anticipate the emergence of Wikipedia and social networking. As for what those other uses are, Jenkins says we will have to wait: "If I had a good sense of other great applications, I would be doing them already."
Build a better baddie
In Jeff Orkin's vision of the future, gamers doing battle in zombie-filled shoot 'em ups will be helping to build richer virtual characters. Orkin worked in the game industry before moving to the Massachusetts Institute of Technology. His system crafts artificial characters by observing how humans play.
The system makes recordings of dialogue and action from games, which are annotated so the software knows what is happening at each point. Then virtual characters in other games can tap into the annotated data to simulate human-like behaviour. Slain by a sarcastic swordsman or glib goblin? Next time, you might have to blame Orkin.
