How to choose the proper actuator for a legged robot? In all mobile machines movement execution is the most important task. If we want to control precisely what our robot does, we have to choose a proper actuator. Robot’s “muscles” are touching the world and giving acceleration, which is necessary to move its body.
Choosing a proper actuator is always a challenge. That challenge appears in all kinds of robots, but it’s especially difficult in legged robots. These machines have many actuated degrees of freedom. All joints must keep desired position, velocity, and torque/force. Additionally, while walking and running, legs cyclically switch between phases of swinging and supporting the robot body, which causes big impact loads. For those reasons, we have to be sure that desired movement computed by the control system is immediately and precisely executed, and the actuator can withstand shock loads.
Good suited actuator for legged robots should be characterized by the following factors:
high torque density — when constructing robots, we often try to do them as compact and dynamic as possible, so the proper choice is actuator, which uses as small volume and mass as it is possible, and generates as much torque as possible. Also, heavy actuator will drastically reduce robot’s dynamics, overall performance and increase energy consumption, what is very important if we are using battery as power supply
low mechanical impedance — it’s not an intuitive quantity, but very important. It tells us about relation about external torques/forces and resulting motor velocity. With high mechanical impedance, the drive is very stiff — is not easy to move the output shaft. External torques/forces are damped. If we use only mass and volume as factors, it will lead us directly to drives, which consist of small motors and gears with high reduction ratios(frequently in range 100:1–300:1). That combination limits overall dynamics as achievable velocities are limited. Also, in case of legged robot gear teeth are often cracking, as during walking we have continuously repeating impacts while touchdowns.
high force transparency — it’s a factor connected with low mechanical impedance. When our drive doesn’t have“obstacles” on the output shaft(like a gearbox, friction, and other resistance to motion) we can “observe” impacts by measuring currents in the drive. All external forces/torques have significant influence on motor velocity and we achieve backdrivability. In case of walking robot, that information can be used for detecting contact between foot and ground without external force/torque sensor
precise motion controller — reliable motion control requires a proper motor controller, which is modulating currents in the most optimal way. In good robotics actuator user can control position, velocity, torque, and often mechanical impedance(spring-damper mode)
high-bandwidth interface — it’s important to get feedback from the drive as frequently as possible, and send desired motion commands too. In legged robots reliable control in real time requires high-bandwidth, low-latency control loops (usually 500–1000Hz)
Now we know the factors, which let us imagine an ideal actuator. It of course doesn’t exist in the real world, so let’s look at available approximations. Below I present some solutions, which show balance between parameters described above.
That leads us to four main categories of drives used for actuating legged robots across the world.
DIRECT DRIVE DD
First, we have direct drives. They usually are just a big diameter, flat BLDC motor. As in motors torque production is proportional to volume, it’s reasonable to keep thickness to diameter ratio low. To properly control the movement, we need a motor controller — when we connect the proper piece of electronics to the motor we get a servo actuator. The advantages are simplicity, low mass, very low mechanical impedance and good reliability. The only thing between motor and environment are bearings, so the force transparency is high. That lets us precisely estimate torques from currents. Also, absence of gears which can be destroyed let us throw a robot from 2 meters safely or kick them. The disadvantage is quite small torque density — all torque production is pure electromagnetic, so motors become hot when the robot is moving dynamically for a long time.
Example of direct drive actuator:
In this case, it’s a drone motor stacked with MAB Drive. In MAB Robotics, we have used DD in one of our prototypes built for Engineering Thesis:
QUASI DIRECT DRIVE (QDD)
Structurally they are very similar to direct drives, but there is an additional component — a gearbox. Usually a planetary gearbox with up to 10:1 reduction is used. That composition of motor with gearbox in this range of ratio has much better torque density than direct drive, as torque is multiplied. Overall dynamics are still good. We have slightly increased mechanical impedance, but we can still detect contacts without external sensors, just by measuring motor current. Drive control electronics are exactly the same as in the previous case — we still have a BLDC motor, so all we need is a three phase inverter with an encoder.
Advantage of that solution is much higher torque density without big mass increase. Force transparency is still acceptable. The only disadvantages are increased drive complexity — we have to connect the motor with the gearbox. It has to be precisely machined and assembled.
QDDs is approach that we are using in our newest prototype:
In this case, the drive is a T-Motor AK80 module with our custom controller.
Quasi-direct drive, due to their medium simplicity and excellent performance in small scale legged robotics, are becoming the most popular actuators.
SERIES ELASTIC ACTUATOR (SEA)
It’s more advanced than QDD. In this case, we have a motor, high ratio gearbox(frequently about 100:1), usually harmonic, and series springs.
A concept construction we can see below:
Photo: Motion Control Lab/http://control.dgist.ac.kr/web/research/rotary-type-series-elastic-actuator/
SEA has excellent torque density, and is able to monitor output torque by measuring torsional spring deflection with encoder. By connecting the motor with spring, mechanical impedance is tuned physically. As we have precise torque sensing, estimation torque from current is no longer necessary. It’s a sophisticated actuator, ideally suited for legged robots, but very difficult to manufacture and very expensive. Off the shelf solutions are for example ANYdrive used in ANYmal B robot from ANYbotics:
Photo: ETH Zürich/https://rsl.ethz.ch/robots-media/actuators/anydrive.html
Last category are hydraulic actuators. Their torque density is the highest. Principle of their work is other than described above solutions — instead of electromagnetic torque, we have extremely high pressure and cylinders connected with precise servo valves and encoders. As torque production is big, and mass small, it’s ideal for dynamic legged robots. The best example is Atlas from Boston Dynamics, which is able to run, jump, and even do a backflip:
Hydraulic actuators also have disadvantages. They are operating at high pressure, so all valves and cables have to be extremely strong. Also they require regular inspections. All malfunctions with pressure in the range of tens, or hundreds of bars are very dangerous for people around. Also, precise control of fluid flow in a closed loop is not an easy task. That’s the reason that technology is very expensive, and probably in the future will be replaced by electric drives at all.
There are a few approaches that let us build high dynamic articulated legged robots. Depending on the budget, project needs and desired performance we can choose solutions, which are optimal for our machine. If you are interested in motor control technology, legged robots or robots in general, feel free to contact us.
Author: Łukasz Antczak
CEO and Co-founder at MAB Robotics