Simple steps


Yamasaki, whose job title was research assistant in the project's symbolic-intelligence group, said he wanted to prove that humanoid robots can walk using motors that are nowhere near as powerful as those used in commercial robots like Asimo and SDR-4X. 

"Using less-powerful motors, I pursued a solution that makes the robot walk with high energy efficiency," he said. He chose motors slightly short in torque to do the job. 

The usual route to making a robot walk is to analyze the human gait and devise real-time control of the robot's joints to mimic it. Honda's and Sony's humanoid robots are controlled in this way. Yamazaki first tried the same approach, "but inexpensive motors do not have enough torque. So I switched to a genetic algorithm," he said. At first, motors with a maximum torque of 25 kg-cm could not make Pino walk, but now, equipped with the genetic algorithm, he steps smartly using 7-kg-cm-torque motors, Yamasaki said. 

In this algorithm, robots "learn" how to walk by themselves through trial and error. For example, large steps in many trials are evaluated and the successful ones "inherited" in the next generation of the walking-control software. "At the very beginning, Pino was just wiggling. But overnight, it had learned by itself to walk [so long as] the evaluating parameter is properly set," Yamazaki said. 

Yamazaki, Matsui and Kitano started the Pino project in October 1999. Pino stood in April 2000 and started walking two months later, in June. That's less than a year from concept to working model. 

A little more than a year after that milestone, in a move timed to coincide with the Humanoid Conference 2001 in November, the team opened up the Pino platform for other developers' use. The name "Pino" and Matsui's Pinocchio-like body design remain registered trademarks, but the rest is up for grabs. "At the beginning, it was just my hope to open the technology," Yamazaki said. "But we [wound up] opening the technology thoroughly, well beyond my expectation." 

Anybody can download source code, revise and redistribute it. Based on the information posted on the Web, developers can build a robot using off-the-shelf motors and CPUs defined in the parts list. Yamasaki said he deliberately used components that can be easily found in parts shops, in the hope that Pino's architecture would be opened. 

Pino's basic architecture is a combination of Futuba Corp. servo modules — 26 in all, falling into three types. The controller consists of two components: an SH2 processor (SH7065) as the master and a complex programmable-logic device (FLEX10K30A) as slave. The SH2 is connected to a PC via RS-232C. 

Frugally redundant


To reduce production costs, the design uses the same components — Pino has 600 components in all — as many times as possible. That leads to some redundancies in the mechanical architecture. But Pino can be built with relatively inexpensive parts. Yamasaki toted up his version's components and materials cost at roughly $15,000.