I'm designing a humanoid robot using pneumatic artificial muscles. I chose these actuator units because

  • they're much cheaper than electric motors of comparable power output capability.
  • instead of needing a high power darlington mosfet array for each motor, only one is needed for the prime movers. Small servo motors individually control the pneumatic valves
  • modern deep learning architectures can "learn" optimal control policies that give reasonable precision
  • pneumatics are way easier to scale to hundreds of muscles than electric motors
  • many pneumatic components can be prototyped with only a 3d printer; custom motor design is graduate school stuff
  • hydraulic fluid (like water) can be used for muscles requiring high stiffness (like fingers)

I'm having a hard time looking past these advantages. Most of the DIY humanoid robot designs I see involve big expensive motors, speed controller, and complex mechanical contraptions. Why haven't cheap (like <$1k) humanoid robots already been commercialized using pneumatic artificial muscles?

  • $\begingroup$ By "penumatic artificial muscle actuated ... robot", I don't mean custom shaped soft robotics. I just mean something like where you have a rigid endoskeleton with joints at the shoulders/elbows/etc. and McGibbon's style muscles providing tension between the structural members $\endgroup$ Jan 31, 2022 at 12:47

3 Answers 3


As cool as they seem, fluidic muscles like these are, unfortunately, not on track to be viable for mobile robotics for the foreseeable future. There are 3 primary issues:

  1. Controls. It's somewhat easy to turn them on or off with solenoid valves, but a humanoid robot needs high precision to walk and manipulate objects beyond just on or off. And it turns out that actuating these fluidic muscles usefully is extremely challenging. Sticking a servo on a valve is a temptingly simple solution but has many issues. You have to direct the high pressure into the muscle at a precise rate to actuate it from the high pressure source and separately exhaust / recycle the low pressure fluid out also at a precise rate to relax it which requires at least 2 valves per muscle or a more complex multiway valve.

This is why after nearly a century of development in fluid controls, the best the industry has come up with are 'proportional pressure control valves' which are very complex and expensive (>$500 per valve, best case) and even then they are slow and inaccurate compared to motors and also hard to miniaturize.

The counterintuitive thing to understand here is that you need to control the pressure to an incredibly high precision to accomplish even the simplest tasks. As the robot interacts with objects and moves the actuators, their volumes and pressures will constantly be spiking and changing and keeping up with this requires a fast and accurate pressure control system. This is compounded by the fact that they are tension only actuators and need to be set up in antagonistic pairs (like human muscles) which requires precise and quick coordination between the opposing muscles.

All that is not to say it's impossible, it's just very expensive and complex.

  1. Cost. It's true that the actuators themselves are cheap tubes but the rest of the system is very expensive. You need a high power compressor to pressurize the working fluid, pressure accumulators to smooth out high demand draws and accumulate expended working fluid to feed the pump, fluid distribution manifolds, pressure sensors, and most expensive of all, the pressure control valves.

The control valves are the killer here, a decade ago Festo built a prototype of what you're talking about called the 'Festo Air Arm' [1] which could even slowly write out large words. But this was realistically nothing more than a demonstration to show off their advanced proportional control valves. I can't find the source anymore but I remember seeing that each valve was ~$2k which seems sensible. No further development was done on this machine though.

On a related note, the Shadow Robot Company makes some of the most advanced humanoid hands available and they used to have a fluidic muscle version available but have since discontinued it because it was too expensive and difficult to control [2] [3]. Their current generation servo based hands are ~300k so it should give some idea of how tricky the pneumatic version was. A recent article about a college that adapted the pneumatic version of the hand put the total price for their pneumatic powered hand system at $350k [4].

Also of note are the pressure sensors. No amount of machine learning can control what it can't measure so you would need a sensor on each muscle which is not only hard to package but also very costly. I suspect >$10k bare minimum in total for a full robot even when mass produced. Trying to control it open loop style solely from the state of the valves would not be feasible either as it would be oblivious to the influence of outside forces acting on the joints. It needs to know if it needs to push harder through something or if it hit something and needs to relax.

  1. Mechanical inefficiency. Regardless of the previous issues, this by itself is basically a dealbreaker for practical mobile robotics applications. Hydraulic systems are generally a little better than pneumatic, but given the large stack of components and high number of moving parts needed to implement these systems the total electrical efficiency is far below electric motors. While brushless motors with gearboxes can achieve >90% efficiency pneumatic systems are only ~10-20%, maybe up to 30% if you have really high quality (aka expensive) parts.

There is actually a Polish group attempting to do exactly what you're thinking of called Clone [5]. They've been working on it for several years now and have had some success building one arm but I'm very wary about their future prospects. If you look closely at their videos, you'll notice that they have only very coarse control of the joints that amounts to basically on or off, I have yet to see any fine controlled motion and I suspect that's due to the reasons I outlined earlier.

On a final note, despite what it may seem on first glance, fluidic muscles like these are actually fundamentally very different than human muscle. As a high level example, if you have an unpowered fluidic muscle robot you can't backdrive the actuators to move its limbs around freely because the pressure is locked up in the actuators. But biological muscle can be moved around without resistance. This points to the fact that fluidic muscles are actually position based actuators while biological muscles are force based. The pressure in the fluidic muscle directly corresponds to a position / length that it wants to be at whereas the chemical power in human muscle corresponds to an output force, regardless of position. This leads to some interesting high level controls tradeoffs and I believe there is good reason evolution chose the force based approach over position based.

All that said, I don't mean to discourage you if you want to pursue this! It's a fun idea and I think the current paradigm of forcibly adapting electric motors to power humanoid robots when they are so different than human muscle is an inherently flawed approach and that there must be a better way. I just think it's important to understand why fluid system engineering and fluidic muscles have been around for over half a century and no company has ever made a viable mobile robotics product with this technology.

[1] https://www.youtube.com/watch?v=2iG1ybuchx0&ab_channel=ThomasPhillips

[2] Current options don't include pneumatic version: https://www.shadowrobot.com/dexterous-hand-series/

[3] Old promotional image at the top shows the pneumatic version in the background behind the motor version: https://en.wikipedia.org/wiki/Shadow_Hand

[4] https://futurism.com/five-fingered-robot-uses-machine-learning-to-perform-dexterity-tasks

[5] https://www.youtube.com/c/AutomatonRobotics

[6] General fluid systems and robotics knowledge: I was the responsible engineer for SpaceX's F9 stage 1 pneumatics system and designed flight critical fluidic actuators / systems.

  • $\begingroup$ I appreciate your write up. Thanks $\endgroup$
    – fionbio
    Jan 26 at 17:06

I would address the concept of pneumatic artificial muscles as being part of what's called soft Soft Robotics, and I would add the compliance they offer as another advantage.

I.e. using a pneumatic hand like the one in the image below, it is possible to hug an object without damaging it. And that hug, is a benefit when operating around humans.

enter image description here

As to why they are not commercialised enough yet, I think its because that they:

  • are difficult to design (the inverse kinematic model, the structure etc), compared to a robot with stiff members and rotating motors as joints. I.e. you have more problems to solve while you are designing.
  • are very difficult to repurpose (i.e. they are usually designed to do a single thing and that is it).
  • offer much less power compared to the alternatives.

However that is not so that they don't already have they niches. I've seen a few applications as exoskeleton for assisted walking to people with kinetically challenged people, which they make ideal candidates due to the compliance they offer to a fitting hug around a body, without damaging it.


figure: (source: Nature)


It is surprising that pneumatic muscles aren't used more often, and I suspect it is because of the difficulty controlling them.

You suggest using servo motors to control valves, but I suspect this technique might be a bit more difficult than you think; the servo must have quite a bit of torque, so will be geared down and quite slow. There will be a relatively long time-delay between the controller initiating an action, and the muscle receiving sufficient air to do that action, making it difficult to do any meaningful control of force, acceleration, speed or position.

I have used Clippard miniature pneumatic valves which are quite fast (5 to 10 msec response time) to try and achieve closed-loop control, but this is still quite challenging, since you are continually balancing the muscle & load forces; the equivalent system using an electric motor can have sufficient torque to make positional control really easy, as the external forces can largely be ignored, if the system isn't overloaded.

  • $\begingroup$ Yes, there is a torque-speed tradeoff, but I think hydraulic/pneumatic actuator demands could be anticipated through a closed loop optimization process. Assuming a model has 1sec lookahead predictions, could something like an N20 geared motor meet the torque and speed demands for bipedal locomotion? $\endgroup$ Feb 3, 2022 at 3:47
  • $\begingroup$ My attempts at doing closed-loop pneumatic positioning were under severe time pressure, and I never really got to grips with the theory, so I can't claim any great expertise in this area. However, I do think it would be a fruitful area for research; as you have said, there are many advantages to pneumatic systems in robotics applications, and I generally found them easier to work with, in comparison with their electrical counterparts. $\endgroup$
    – jayben
    Feb 3, 2022 at 13:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.