Best platform for self-balancing and running biped robot?

Community General Discussions
21

Another thing I though after

Another thing I though after reading this complete thread. The one who would like to have a biped and make a movement mechaning should forget about any lasers, AI and any other sensors except IMU (and maybe a load cell). It would simplify the task drastically. Just make some nice walking biped controlled with a PlayStation gamepad, and the work will be considered 100% done. The motion and balancing should be done the same way as it is done in current self-balancing Segway-like robots. Except this one will be on legs. I see lot’s of options and already found out an algorythm for the job:

  1. Inverted pendulum of 2 sticks one on another on a platform - you rotate the platform, the construction stays vertically - possible to do even with Arduino and couple of motors. Thoughs: Do not use servos as they seem to be too inertial with their cheap plastic gear sets, only stepper motors with an additional sensor to track rotation.
  2. 2 sticks set of 2 sticks on a plaftorm with another platform on top - you rotate a platform, two sticks should hold vertically a pelvis - top platform
  3. 2 sticks sets on 2 separate platforms, you move one platform up/down, the other sticks and the upper platform adapts - I believe Arduino is still enough here with 4 additional motor controllers which will control about 7 motors. 
  4. One stick moves itself up and down, while the other adapts. 
  5. … to be continued, this is enough for a year of work I believe… and if anyone does this and show results, I would be happy to work with you on the task and contribute what I can.

Another thoughs - the system should not know their center of mass. I’ve actually found a research here and find it to be too “smart” for the task, as the task should self-balance and react on the IMU, nothing to “calculate” actually here except of rotation/position of motors/leg.

What do you think?

22

This robot walks properly,

This robot walks properly, with a real human gait:

https://www.youtube.com/watch?v=-Vg-BdXps50

My jaw dropped when I first saw that. And it’s made by a hobbyist, not some large research lab or company. Admitedly, the small scale makes it much easier, becauser the servos can be relatively stronger and the inertia is smaller, but still it is amazing. Also watch other videos by that user.

23

Asimo keeps balance by

Asimo keeps balance by relying on the force sensors in its feet, which in turn let it know its Zero Movement Point, which is a kind of generalized Center of Mass, adjusted for inertia. It can then do motion planning for the feet so that the ZMP never leaves the Area of Support (the polygon you get by drawing around all legs on the floor), which guarantees that it is stable. But in order to make that planning easier and simplify the calculations, it walks with bent knees, keeping its center of mass at constant height – that makes the gait look unnatural and makes it inefficient energetically (takes about 20× more energy than human walking). But it works well enough to put into a robot that is actually sold as customer product.

Generally, the problem is much more problematic than inverted pendulum, because you only get short moments when the foottouches the ground when you actually have traction and can do anything. Unless, of course, you go the typical route of hobby humanoid robots with humongous feet and ankle servos strong enough to lift the whole robot. But that’s not how humans walk – we could as well walk on stilts with no feet at all, and it would be a natural gait. So you have to really plan ahead, including inertia in your calculations, and then you have to be able to modify that plan with information from sensors to keep your balance. And that’s a set of nasty non-linear equations without a closed-form solution, which you have to solve numerically using iterative algorithms. We can actually solve them. With current computers, calculations for motion planning for 2 minutes of walking take about 12 hours to compute. Unless you cheat and simplify like Asimo does.

 

24

Power and Weight

It’s not computing power that holds this back.

I would like to see a  humanoid walk a normal walk at less than 1fps, and with steps of a couple of feet or so, not baby steps. If that is possible then you can use any servos that have sufficient torque to move it.

Otherwise this turns into a much more complex calculation of how much power is needed to drive it to a certain speed .  That depends on the geometry.

I have done these calculations for quadrupeds as I making one based on the architecture of the cat and about cat size. With hobby servos in the 15 to 20 kg-cm range I do not see this moving faster than 1fps, nowhere near fast enough for anything but an amble.

For a humanoid of a few feet tall this is a daunting challenge. I would look at these:

https://www.servocity.com/html/heavy_duty_linear_servo__25__l.html#.Vfog1RFViko

As far as processing, I can tell you what is sufficient for the mechanics. I’m using an Arduino DUE,and that is enough to handle the floating point I am feeding it. The servos run off this:

http://www.hobbytronics.co.uk/pwm-servo

So, how big? how heavy? What will it look like? How will you rotate the joints?


 

 

 

 

 

25

Now, this is a concrete

Now, this is a concrete feedback, thanks. My intention is first to make it walking on it’s place, just raising one leg and balancing on the other and then the same with another leg. Then by adapting balance of an “upper body” start making baby steps into sides. As I said, I believe it might be possible to do within a year or so (with my current level of knowledge) and I am mostly curious on what kind of showstoppers I fail.

At the moment I don’t like servos and decided to use a stepper with an optical encoder. I also think that the body size/weight will be calculated from the available motors. I start with the smallest motor and will see if I have enough torque with it. Another very important thing is a joint connection. It must be something smart - btw, how do joints work on industrial arms? I’ve seen a fancy oval gear-set, forgot the name of it.

I have already modeled and ordered a samle for shapeways, hopefully it works, but in general there should be a special durable low resistance plastic for a joint etc. So things are interesting overall.

I don’t think at the moment about overall mass - it’s too early, I need to make 2 sticks balanced ;-).

26

I have ordered yesterday 2

Wow, the explanation of Asimo walking is great, I was always curious about why he walks so ugly… :slight_smile: But the approach is just correct, I like. :slight_smile:

I have ordered yesterday 2 motors and 2 optical encoders for them. Will go to my step 1 next weeks (when I hopefully receive Adafruit motor controller, MPU5060s and a proper table-mount for the system). We’ll see and I will update you with results, really curious what the showstoppers can be there. What you describe, seem to be similar to this, but as I mentioned before, I believe there is no need in such heavy computations. Do you really believe that moving 2 motors (to move the pelvis left-right) requires calculations of center of gravity? For some reason I think you can do this by just reacting on IMU rotation. 

As for the “force sensors” in the feet, I thought about it before and considered adding some load cells in the “feet” later. But my intention was just to determine if the “foot” is on the floor or not, thus “enable” and “disable” the leg.

Overall I believe there should be a compromise between mechanical movement and balancing computation, and by writing a library for this hybrid approach one could get these legs be not only mechanical but adapting.

27

btw, this might be a right

btw, this might be a right servo… http://harmonicdrive.net/products/actuators/rsf-supermini/

hm… really fancy stuff: https://www.youtube.com/watch?v=AdpyLo-tjsE ,

but can’t find how much it costs…

28

I really suggest that you

I really suggest that you watch the MIT “Underactuated Robotics” course, where they explain this and more, and all the math involved. You can also enroll for that free course online. Here are the lectures from 6 years ago: https://www.youtube.com/results?q=mit+underactuated+robotics+course

 

Update: Here’s the link for the course. It’s free, and I really recommend it to anyone building robots. https://www.edx.org/course/underactuated-robotics-mitx-6-832x

29

Thank you very much for the

Thank you very much for the links! :slight_smile: I will gladly check it of course.

30

arisetobe - i had/still have

arisetobe - i had/still have a similar goal/dream… to make a dynamic balanced biped humanoid with a true non-static human-like gait… largely inspired by DrGuero2001 & Boston Dynamics ATLAS

I think it is possible… so far, have only achieved static-balanced gait, and the keyframes are set manually (in a custom excel sheet which takes in servo offset values from center in a large time-varying table, and outputs a matrix dataset of servo values and keyframe timestamps) with arduino performing iteration between frames only, no onboard kinematics: youtube - robot walk gait 2 

He does have an onboard IMU (MPU6050) and yaw-pitch-roll info is available to the servo control board, polling at about 100Hz - currently this data hasn’t been integrated into control algorith… starting point will be to just adjust ankle/hip servo positions to remain statically upright on a dynamically shifting floor in both pitch/roll. Later, and infinitely more complicated, would be to incorporate IMU data into dynamic walking. First though i’ll likely need a complete rewrite so walk gait is driven by IK instead of current method.

Have done some basic preliminary work on using assessment of incoming MPU6050 data to detect an external force, in order to kind of cheat and allow pseudo-flexible/compliant joints in order to accomodate disturbances in balance - servo joint compliance test 2 - also have 4 x Force Sense Resistor pads to integrate onto the feet, to receive feedback on which foot is currently touching ground, and whether weight is shifted forward/backward. 

So, much work yet to be done

If you have access to a 3D printer - the frame i’ve based my bot on http://www.thingiverse.com/thing:786036 - whilst not perfect, is at least a quick way to get up & running for basic testing.

31

This is comment to osh, as

This is comment to osh, as it is better to maintain new comments on top and not within pages.

I haven’t seen Dr.Guero other versions than driving a bicycle. Now this: https://www.youtube.com/watch?v=_2cfc75M8Os and this: https://www.youtube.com/watch?v=6YPsLJ0jwDI is actually almost exactly what I am talking about, and your way of thinking I find also right. And it shows that what we are talking about is possible to do by ourselves and that is even with simple small servos it will work, as there is such a great proof of concept.

Few comments though: 

1. Construction - I strongly believe that there should be something better than just servos and these simple joints. I want bearings and precise gears. I want to eliminate any tilting on joints, thus making the whole construction more stable and less noisy. For example something like this but small: http://www.pic-design.com/bevel-gears.html

2. Of course motors. I am still going to experiment with proper steppers and geared DC motors until I have to try servos. Can anyone suggest a proper servos for this kind of project?

3. Gaits and mechanical walking - your excel export is cool by it’s nature. And a correct approach, but I would suggest to export it directly from an animation editor. Making one gait that would look “normal” (and human beings are very good at understanding of what “normal is”, that is why it takes some time to animate a gait cycle - btw, Disney animation tutorials explain a lot about what “normal” gait is, and their gaits, despite being a bit exaggerated, nevertheless physically almost correct). 

4. Inverse Kinematics - I believe it might help a lot with programming of hands positions, but not sure if it is necessary on the gait. You see, from my experience, to make a “normal gait” (see previous point), it requires not only IK but also FK to adapt it (and hell amount of keying - you won’t program it, it should be either made manually or motion captured - or better both). And once mechanical part is taken from animation editor, the task of the balance is just to adapt. Thus, I see no need in using IK in adaptation.

5. Accelerometers. It might be possible to use something better than MPU5060 (maybe), as I have heard that MPU5060 is noisy and requires additional filtering etc. Though together with MPU5060 I have already ordered LSM9DS0 and BNO055 and already tried CMPS11 (which worked surprisingly good btw)… Maybe others, specifically BNO055 should be better, will see.

And because there are too much work to do for a one person, to make the job more efficient, I suggest you to join the “team”… :slight_smile: https://www.robotshop.com/letsmakerobots/lets-make-a-biped-together

…I am a professional project manager btw, working with IT projects for more than 10 years (several years run a game development company, while we developed our own engine from scratch), so I know how to manage R&D projects and our planning and approach will be top notch. (now I feel myself as a recruiter though, and immediately remembered uncle Sam’s “I want you for U.S. army”… :slight_smile: )