Note: there are currently 7 HUBO 2s in the USA
At ICRA 2012 Dr. Jun Ho Oh (KAIST, Hubo Lab) gave a revealing presentation on the development of his full-size humanoid robots. The latest, HUBO 2, represents 10 years of research and is now commercially available to anyone with $400,000 to spare (around a dozen units have already been sold to universities in the United States, Singapore, China, and South Korea). It’s great to sneak a peek inside his lab, when so many other humanoid projects are so secretive. It can be viewed in its entirety (about an hour) courtesy of TechTalk.tv, but we thought we’d break it down into some of the more exciting bits for you here.
Dr. Oh states up front that his work is to build the hardware, not the software, of humanoid robots. This somewhat lopsided approach is something that many Western pundits have criticized, but to me it only seems natural (if not essential) if you plan to build a highly competent platform like HUBO 2. Software, such as speech recognition, can be quickly added to the robot’s repertoire since (for the time being) it offloads higher-level processes to external computers. He begins with an overview of DARPA’s recently announced robotics challenge (perhaps suggesting HUBO 2 could participate) before briefly covering its evolution stemming from the KHR-1, KHR-2, Albert HUBO, and HUBO FX-1.
(It should be noted that, contrary to Dr. Oh’s offhand remark, the KHR-1 was not the first full-sized walking humanoid built outside of Japan: China’s Xianxingzhe (2000) preceded it, and Technical University of Munich’s JOHNNIE was developed concurrently – there may also be others that aren’t coming to mind)
The locations of the elbow and knee joints are noteworthy, having been offset by a few more degrees than is typical. They look slightly odd as a result, but their location allows the limbs to bend almost as much as their elastic human ancestors. This flexibility means HUBO 2 can bend down and pick up objects from the floor without much difficulty. It is also revealed that the legs have multiple motors per joint for added power.
A clip of HUBO 2 doing push-ups (although missing its delicate hands and fingers) proves it is no slouch. The hands, attached to a 3-DOF wrist are particularly remarkabale. They weigh less than a pound, contain hundreds of components, and have a grip strength of about 7 kg. That’s enough static strength to open a car door, operate a gear shift, and hold heavy power tools. Its individually-actuated fingers (only recently implemented by Honda’s All-new ASIMO) can automatically adapt to the shape of whatever object it grasps.
(Often humanoids aren’t equipped with such sophisticated hands; examples developed for DLR’s Justin and NASA’s Robonaut 2 exceed those of HUBO 2)
Video (Hand grip tests, Battery stress test):
Another fun surprise was a test of HUBO 2’s LiPo batteries. The robot walked steadily on a treadmill for two hours, covering approximately 2 km (1.24 miles), before its batteries finally died. Later in the presentation the robot’s energy efficiency was improved significantly by switching to the so-called “straight knee” walking gait, suggesting it could cover even more ground in a single charge. In a related note, Dr. Oh revealed that HUBO 2’s running speed has been increased to about 4 kph fom the 3kph it achieved earlier.
Unfortunately missing from the presentation is HUBO Lab’s quadruped, a fairly secretive project no doubt inspired by Boston Dynamics’ BigDog. This is something I really want to see in action, if only to compare the two. Dr. Oh, like most roboticists around the world, has been paying close attention to Boston Dynamics’ latest projects. Even HUBO 2 performs some PETMAN-inspired actions to test its full-body dynamic control.
Video (Push-ups, Bending, Stretched-leg walking, Whole-body motion control):
However, some of the lab’s other projects were shown. First was the HUBO Arm, which has longer links (like the Barrett WAM arms), providing a greater reach. The arm and hand are extremely light-weight and agile. Next is the HuboQ, a Quick Delivery Robot, which is essentially a Segway with one of HUBO 2’s legs mounted to the front of it. It zooms around at 12 kph (limited from a possible 40 kph for safety reasons), carrying up to 100 kg. Proferring hors d’oeuvres, it was a big hit at a recent event. What’s more the HuboQ can carry a passenger, making for what Dr. Oh described as a “very interesting toy in my lab”. I’ll bet!
When asked what to expect from the next generation of the robot, he dodged the question. Mechanically-speaking, HUBO 2 stands shoulder-to-shoulder with the most advanced humanoid robots in the world that are driven by electric motors. It may not be able to run as fast as ASIMO, but is impressive in many other ways. HUBO 2’s weight, for example, can only be reduced by about 5~6 kg – marginal savings that would only be possible by switching to a carbon fiber frame and exoskeleton. It may also be that Dr. Oh expects the next steps to be taken by his many students, which have now become professors in the own right.
Like Honda and AIST, KAIST’s researchers seem to have reached the current technological ceiling, having achieved only incremental gains over the past few years. This would explain why Honda was so slow to unveil the latest version of ASIMO, and why the HRP series appears to be stagnating (albeit at a very high level of sophistication).
PhD Students who have helped create HUBO
In terms of making changes without completely switching paradigms (to artificial muscles or hydraulic actuators), I would argue that there is much to be done in the areas of compliance and tactile skin sensors, but I guess these topics aren’t too much of a concern at KAIST. Skin is something I’d like to see more of on robots, but currently there are only a few institutions that seem to be working on it (Italian Institute of Technology, Tokyo University, and Osaka University).
Furthermore, recently JSK Lab’s HRP3L-JSK showed us that significant performance enhancements become possible through the use of high-power capacitor-driven motors. Still, it probably wouldn’t hurt to shift the focus to reducing costs and improving overall robustness and artificial intelligence, and that’s what is bound to happen now that HUBO 2 has been adopted by multiple labs.
If you have the time, it’s worth watching the presentation in full.
All images and videos copyright KAIST Hubo Lab | Erick Oh