Update: I have launched a GoFundMe campaign in order to fund the immediate next step of the project which is the muscle survey. The faster this is funded the more time I can spend on getting the results needed.


It’s been a while since I last posted here. This has been due to the usual case of when I have resources I don’t have time and when I have time I don’t have resources! Also there were fundamental problems with how I and a lot of other people had been approaching the concept of humanoid robotics. It’s taken time and looking deeply at biology to understand the problems.

So what is the fundamental problem when it comes to building a humanoid robot?

I believe the answer is the actuators. Most of the time robotics use electric motors, either in custom housing or commercial servos. Almost every robot from hobby kits up to research androids use electric motors. Every humanoid robot has been designed around the choice of actuator, and their issues have led to compromises. The more I learn about how the body works the more restrictive these compromises are.

So what is the solution?

The solution is soft robotics. Pneumatics or hydraulic actuators made from soft materials. Pneumatic Artificial Muscles (PAMs) have a long history and have been used for basic humanoid inspired robotics for decades.


Why haven’t they caught on if they’ve been tried so many times?

PAMs have a myriad of their own problems. Firstly there are many different types, ranging from McKibben muscles to Pleated to PEANO designs. Generally speaking a research group comes up with a design, runs some tests and makes a behaviour model and then… Only the Mckibben muscle seems to be in relatively common usage.

Mckibben Muscle gif. Author Rocketmagnet (

PAMs are notoriously non-linear and can only provide force in one direction. The forces they can produce do not give them any benefits over conventional pneumatics, and due to sliding and flexible parts they often have major wear and tear issues.


PAMs also require valves and pressurised gas to operate on. These are optimised for general commercial usage and not for a humanoid robots and so tend to be heavy and bulky. However there are many ways to make a valve and so it is not impossible that more effective ways haven’t been used. One example would be to fit the valves and manifolds (shared gas chambers) into the bone like structural elements.

Robot design with many bulky solenoid valves.

There has also been a structural problem with most PAMs. There are often strictly linear and mimic the fusiform type of muscle. This type of muscle has a single anchor at both ends and the muscle fibres run straight from one to the other. This is the classic type of muscle that people imagine when they think muscle. However there are very few muscles in the body that are actually structured like this.


A range of different muscle geometries.

In skeletal muscles the force generated by a muscle is proportional to the physiological cross section. In a fusiform muscle this is the cross section you get if you slice the muscle in half at it’s thickest point. This results in muscles ‘geared’ to large strain and low force. But there are other muscle geometries, such as Pennate muscles. These muscles have their muscle fibres arranged at an angle to the tendons and the direction of force. A number of muscles that help flex the fingers are of this type. This generates a larger force but sacrifices strain. The fingers can only bend so far and are needed to carry heavy objects, so this is fine.

Cross sectional area

So far PAMs have mostly ignored pennate muscles, and up until recently haven’t occurred in geometries other than linear cylinders.

Another issue for PAMs with that whilst they are not difficult to make individually they have not been built in forms that are amenable to mass production. Hence making manufacture of a humanoid a very labour intensive process.


Another issue that has held back humanoid robotics is a misunderstanding of how muscles work together. Everyone is taught in school about antagonistic muscle groups, but what if I told you that arguably there are no antagonistic groups and there are common cases where such ‘opposing’ muscles contract together. An example is the Lombards Paradox, where the hamstrings and quadriceps contract at the same time whenever you go from a sitting position to a standing one. This is possible due to the precise geometry of the muscles including the action of bi-articulated muscles (muscles that span more than one joint).


If I was to describe the approach taken by evolution in designing skeletal muscle systems I would describe 2 main groups of skeletal muscles. These would be that single joint spanning muscles have evolved to provide precise joint control and the muscles that span multiple joints that provide power.


Single joint spanning muscles often lie close to the joint and are quite small in size (eg rotator cuff muscles). They are often numerous and are designed to stabilise joints. The more degrees of freedom and joint can have the more of these muscles to keep them under control. These are actually the muscles that are replicated in many humanoid and industrial robots.

Rotator cuff muscles

Compare the joints of these humanoid robots and actual human joints.

The multiple joint spanning muscles are often more superficial to the single joint spanning ones. This is because they need freedom to slide into new positions as the joints move underneath them. These muscles are often large and can provide a large amount of force. Any robotics system using these many joint spanning muscles is at an experimental stage only.


Some other aspects worth mentioning about how muscles group together. Evolution has optimised the positioning of muscle mass around joints, based on where the forces are actually needed. A human engineer would be tempted to provide an equal amount of torque in all directions, even when a) gravity does the job for you and b) there’s only so much torque required in any particular direction.


The direction of force generated by the muscle is also affected by geometry. A muscle can apply a torque in one direction around a joint, but when the joint goes past a certain point the torque is now in the other direction. The muscle slips past the centre of rotation and can now provide more force in the direction the limb is moving. This can be seen in several muscles in the shoulder. This is a brilliant design, as it allows the majority of a muscle group to contribute to the force in one direction. This in turn lowers the overall amount of muscle mass required as a single muscle can fulfil more than one role.


So I hope this has been eye opening and helped you think of muscles in a new way.


What is my next step?

I want to create a design family of PAM that fulfils a number of criteria.

  • It’s not limited to cylindrical geometry.
  • It’s not limited to fusiform geometry and can be designed to create uni and multi pennate, as well as plannar and converging geometries.
  • It’s easy to manufacture.
  • Can manufacture multiple muscles in one go, and can even include many muscles in a single structure.
  • The qualities of the muscle can be predicted fairly reliably.
  • The muscles can be created at a range of scales, from 1cm in size up to tens of cms.
  • Can be anchored to bone like structural elements of the robot.
  • Can contain non actuated regions in order to mimic tendons and fascia.


What do I need to do this?

I will need a reasonably anatomically correct life size skeleton. Preferably something mass manufactured and cheap, so that I can make changes to the skeleton without worry that this is the only one I have.


I will need to measure how long the muscles will be at rest, full contraction and full extension. This will allow me to define parameters for the muscles.


I will need to experiment with PAM designs such as PEANO and flat PAM. Tests will involve predictability, form factor and ease of manufacture. For this I will need the equipment to make the muscles, solenoid valves and electronics. I will also need something like a Universal testing machine in order to characterise the muscles.


Future investigation will involve making the additional aspects of pneumatics more suitable, eg incorporating pressure vessels and valves into the ‘hard’ structures of the body. The PAM may plug directly into the valve output coming out of the bone.


Why do all this?

Beyond the cool factor humanoids have evolved to work well in a range of environments and a range of motions. Sure a specifically designed robot can do a particular job better than a person can but then it can’t do another 99% of what a human can do.


Telepresence would also be a huge use. Having a humanoid form with human movement and dexterity will allow a human to inhabit another environment with much more ease. A humanoid robot could dive down much further than a human can and be unencumbered by oxygen tanks and other life preservation equipment. The same could be used in space, rather than perform an EVA. A humanoid robot geologist on the surface of mars could be controlled by someone orbiting above or on one of the martian moons. Pneumatics or hydraulics would require relatively few changes to make a robot work in these environments. Mostly tweaking materials properties based on temperature changes.


This work would also be applicable to elements such as prosthetics. Heavier elements in the pneumatic system could be moved to a pouch or back pack, leaving a lightweight but strong prosthetic.


Pneumatics can also be run on various power sources such as compressed CO2, or even small internal combustion engines. These have comparable or better energy densities to modern batteries.


Whilst this has been focused on humanoid robotics, there is no reason why it couldn’t be applied to a range of animals or fantasy creatures. This would be a boon to animatronics as a source of non-jerky lifelike movement, with greater complexity than can be achieved using servos and wires.


I hope that this fleshes out a little on the steps I want to take in the coming months. There are a lot more details than are mentioned in this post. I’ve not mentioned other structural components like fascia, or how the plan to make the muscles involves sacrificial 3D printable molds. Susan from is very kindly printing some PLA and PVA molds for me, which I hope to receive in the next week and so I’ll begin a systematic process of building these designs.

So here’s hoping that I’ll have a new post and perhaps even video for you in the next couple of weeks.

Concept render of how the muscles may be arranged.