Robot unravels mystery of walking

News

Not a member yet? Register for full benefits!

In July 2007, the University of Gottingen, in Germany unveiled the product of four scientists; a robot that walks like a human.

RunBot is a self-learning, dynamic robot, which has been built around the theories of Russian neurophysiologist Nikolai Bernstein.

In the first half of the 20th century, Nikolai formulated most if not all of the key issues to be faced in recreating modern co-ordination of movement. These included the degrees of freedom problem, motor equivalence, and non-univocality of motor commands and peripheral effects. His book, "Co-ordination and Regulation of Movements", published in 1967, has been used for 50 years in the study of human motion. Now, for the first time, it has been applied to robotic motion.

Nikolai postulated the then-unthinkable notion that any kind of human body movement results from an infinite variety of possible combinations, or degrees of freedom, of neuromuscular and skeletal elements.

He believed the system was thus self-organising, and by and large operating without constant supervision from the central nervous system. That is, save when constraints of the body changed (a child growing up, onset of osteoporosis or arthritis) or when the conditions outside changed, (such as a floor giving way, or suddenly encountering a fluid spill, change of gravity, or being pushed from behind).

In robots, that has not been how the process has been achieved until now. Usually what happens, with all robotic walkers, from the humblest spider-bot to the big guns like Asimo, is they are kinematic walkers. In other words, they and calculate every single angle, every millisecond, for every possible location of the foot, the leg, the knee; every joint. The brain at the centre of the robot is performing hundreds of thousands of calculations, and the whole thing is very precise, very deliberate, with no margin for error. This is why it has taken closing on twenty years to teach Asimo to climb stairs - every possible contingency must be factored in, and the storage and processing required is truly vast.

Kinematic walkers are slow. They handle walking as a massive engineering hurdle and so stomp around with very precise steps - everything equal and measured; devoid of all personality. If the terrain changes, they pause, analyse the situation, consult the database for new angles, new procedures to try. If nothing is in the database, they cannot go any further, they simply do not understand.

Using Nikolai Bernstein's theories of locomotion, RunBot was created, as a passive walker. Unlike kinematic walkers, passive walkers do not even attempt to calculate every possible angle. Instead, all they are fussed with is keeping upright. What is sticking out where, is really of no concern. All is good, so long as they remain upright. Each leg looks after itself. Processors in the leg itself deal with most situations. It hurries along, putting the segments in place according to a simple mechanical physiology. Best thought of as a kind of inbuilt instinct, the processors know that the step must spring at the right time, to push the robot forward into the next one.

Passive walkers can walk on level ground, they can walk down a slope unaided, propelled by gravity and kept upright and moving through this mechanical physiology.

Learning from its mistakes

Rather than have a huge database of every possible angle set for every possible situation, RunBot was also built with a neural net. It makes mistakes, frequently - a misstep limb lands it on its rump. However, each time that happens, it learns. It learns what angles do and do not work on which situations, and it trains its legs. A central processor is involved, but only in training the limbs how to move. It is not involved in individual steps; only analyses and makes decisions when something has gone wrong. In this way, the robot rapidly learns.

It also eases control. Rather than the central robot brain having to do everything itself, with this new breed of passive walkers, the central brain is freed up. It does not have to think about walking, it just walks until the central brain tells its legs to stop, or change direction - exactly the same way as a human brain does.

As it is so similar to how humans walk, it should come as no surprise that RunBot can move at speeds of more than three leg lengths per second, slightly slower than the fastest walking human - and is by far the fastest walking robot presently in existence, far, far faster than the kinematic walkers.

Applications in Prosthetics

The application of a passive walker utilising this kind of technology is easy to envisage in prosthetics. The hard part would be integrating it in. As they are modelled on what all the evidence points towards being the same way that human legs operate, then the benefits of a passive movement limb with its own integrated circuitry, keeping it moving as to the last command until it fatigues (not going to happen with a robotic limb) or until it is told to stop, are crystal clear.

The problem comes in how to integrate such straight into the central nervous system. At present, we are nowhere close to untangling all the signals the brain sends down - we have a mere handful, not enough to walk on. Over time, we will gain more, and at an ever quickening pace.

It is likely to envisage passive moving prosthetic limbs, within two to three decades, and this is most definitely an avenue of both interest and promise.

Applications in Virtual Reality

Applications in VR of this technology do not come to mind so readily, unless you consider the issue of avatar movement. At present, most 3D avatars are pre-animated things, whose legs swing when they are walking, with no relevance to the surface they are walking on, or even if there is a surface under them.

As current technology stands, we have no requirement for anything more. But, as telehaptic systems continue to improve, there will come a point, rather soon, when communal VR simulations are sufficiently advanced enough to feel the positions of the limbs of your avatar

At this point, likely within one to two decades, a model of walking for avatars so they can place feet properly upon every surface they encounter will be required. Kinematic walkers will likely be the first to spring to mind, as they are the mainstay of robotics at the moment. However, their immense computational cost is likely to consign them to a spot 10-15 years further down the road for most applications.

A passive walking model is simple enough to be executed on the same computer interface as the user, and just copy the movements across to the server. Any VR world that implements this technology is very likely to have sizeable marketing leverage over the competition - as well as providing a much higher fidelity experience to its userbase.