Armed with Watson smarts, Pepper aspires to be a robot salesman
Japanese robot Pepper is getting an intelligence upgrade via IBM’s Watson, but that doesn’t make interacting with the real world any less challenging.
The humanoid will channel Watson’s artificial intelligence (AI) knowledge base when it starts work as a sales clerk next year at Yamada Denki, a major electronics retailer.
Pepper has helped sell goods such as smartphones and coffee machines before, but engineers hope Watson’s ability to suggest relevant information will lead to richer interactions with customers. The robot can be programmed with a specific goal in mind—to move stock out the door.
Laden with sensors, the cloud-connected Pepper has been marketed as a communication robot with a feel-good vibe and penchant for jokes, but developers are trying to make it more useful. One problem is the bottlenecks inherent in “embodying” the Watson AI platform and its natural-language processing powers in the real world via a robot like Pepper.
In a demo on Thursday, IBM described how Pepper can identify customers in a shop, approach them and strike up a conversation, and answer their questions about products such as flatscreen TVs with an eye to making a sale. It could even bring related information, such as the start of 4K broadcasts, to bear. The demo took place at a tech show in Tokyo hosted by mobile carrier SoftBank, which put Pepper on the Japanese market last month.
Watson servers in the U.S. had been fed information related to conversations about TVs so that the software could choose the best answer when given a query. It employs a dialog manager, a question-and-answer engine and speech-text interfaces.
But amid the noise of the show, the robot often failed to recognize what was being said by an IBM staffer acting as a customer. That doesn’t bode well for working at Japanese electronics retail outlets, which resound with shouts from human clerks trying to get attention.
Pepper also had difficulties in another demo designed to showcase its ability to recognize everyday objects. Staff from SoftBank cloud service firm Cocoro SB, which engineered the robot’s “emotion engine,” presented things like candy, a bar of soap and a tube of toothpaste to a camera on Pepper, and named each in turn.
Later, when presented with the candy, for instance, Pepper said in high-pitched Japanese, “Let’s see….that’s a bar of Kao White soap.” Onlookers burst out laughing.
The firm attributed the goofs to a lack of training data in the neural network the robot uses to identify things. The network compares the features of an image to high-level descriptions of other images in its database to find a match.
As Pepper practiced more throughout the day, its performance improved. But when shown a voice recorder, which was something totally new, it identified it as a tube of toothpaste. After being presented with the recorder in three different orientations, it was able to identify it correctly.
That sort of step-by-step knowledge building is essential for recognizing objects in its environment.
“It has no broad understanding of things. It just looks for small patterns,” said Cocoro SB engineer Akira Takanami.
Watson could significantly enhance Pepper’s knowledge of new items. When introduced to chocolate, for example, Pepper could draw upon the platform’s knowledge of the food and start talking about chocolate varieties and recipes.
“With Watson, it can answer questions more deeply,” said Shu Shimizu, a senior manager for cognitive computing at IBM Japan.
Robots could be better at selling than some human counterparts, at least if one experiment is anything to judge by. Scientists from Osaka University’s Intelligent Robotics Laboratory set up a lifelike female android as a clerk in a department store and had it try to sell US$100 cashmere sweaters. During the experiment the robot dealt with twice as many customers as its human counterpart.
SoftBank said it will start accepting lease orders for Pepper from enterprises in October. The robot will be able to approach customers, conduct questionnaires and act as a receptionist. The bot’s monthly “wage” is ¥55,000 ($444), which is less than half that of someone working at Japan’s average minimum wage.
“Pepper can work 24 hours a day, it won’t complain and it will never be late,” SoftBank CEO Masayoshi Son said.