How Robotic Exoskeletons and Exosuits Work: Sensors, Control and Safety

The short answer

A powered exoskeleton is a wearable robot that adds force to a person’s joints. Sensors measure how the wearer is moving, software works out what phase of a movement they are in and how much help to give, and motors deliver that help through a frame or through straps and cables. The person stays inside the control loop the whole time. [1]

There is no single way to do this. Exoskeletons for paralyzed users, soft suits for people recovering from a stroke, and passive back supports for warehouse workers use different hardware, different controllers and different safety rules. The review by Baud and colleagues describes “a considerable number of different controllers” in the research literature. [1]

A simple picture: helping someone push a heavy door

Imagine helping a friend push open a heavy door. You need to push when they push, in the same direction, and roughly as hard as they need. Push too early and they stumble forward. Push too late and you add nothing. Push in the wrong direction and you fight them.

An exoskeleton faces the same problem, many times per step. The comparison has a limit: you can see your friend’s face and hear them, but the device cannot read the wearer’s mind. It only has its sensors, so it estimates what the wearer is trying to do from joint angles, foot pressure, body motion and sometimes muscle signals, and it can be wrong.

Rigid exoskeletons and soft exosuits

[2] [3]

The Harvard exosuit in the 2017 stroke study was worn on the weaker leg, anchored by textile bands at the waist and calf. One cable in front of the ankle lifted the foot during swing, which helps the toes clear the ground. One behind it helped the ankle push off during stance. On the treadmill the actuator sat off the body; for walking overground, a 2.63 kg actuator and a 0.56 kg battery rode on a waist belt. [3]

Rehabilitation, mobility and industrial devices

Medical exoskeletons are regulated as medical devices. In the United States, the rule for powered lower-extremity exoskeletons requires clinical testing that records adverse events, software verification and hazard analysis, electrical, battery and electromagnetic testing, a training program for clinicians, users and companions, and labeling that warns against risky activities and environments. [2]

Industrial exoskeletons support workers rather than treat a condition. NIOSH distinguishes passive devices, which use springs and counterweights, from active ones powered by motors, pneumatics or hydraulics, and lists back, shoulder and arm, tool-holding and leg-assist types. It also notes that most studies so far have involved small numbers of participants, mostly in laboratories. [4]

The international personal care robot standard ISO 13482 covers physical assistant robots, but its scope excludes robots used as medical devices. A rehabilitation exoskeleton and a factory support suit can look similar and still fall under different rules. [5]

How the parts interact

A wearable robot is a loop with a person in it. The wearer moves. Sensors measure the movement. Software estimates the situation and decides on an assistance command. A low-level controller drives the motor to produce that command. The force travels through a gearbox, a frame or cables, and cuffs or straps into the wearer’s body, which changes the next movement. [1]

Baud and colleagues organize controllers in three levels. The high level selects the overall mode, such as standing, walking or climbing, and mode changes are usually separated by at least several seconds. The mid level runs every control cycle: it detects the gait phase and computes the desired joint torque or position. The low level makes the actuator produce that output, using position or torque controllers. The authors note that most research effort goes into the mid level. [1]

What the sensors measure

[1] [3] [6] [7]

The 2017 exosuit combined several of these: gyroscopes on each shoe split each step into sub-phases in real time, load cells in the textile measured the delivered force, and potentiometers measured cable position. [3]

Estimating what the wearer is trying to do

Intention cannot be measured directly. Controllers infer it from patterns: a heel strike marks the start of stance, a rising shank angular velocity suggests the leg is swinging forward, a lean of the trunk may signal the start of a step. For people with paraplegia, Baud and colleagues note that IMU data often comes from the upper body, because the legs themselves cannot start the movement. [1]

Muscle signals are attractive because they appear before the movement does. Fu and colleagues report that EMG onset can be detected about 50 to 100 ms before physical motion. The same chapter lists the problems that make EMG hard to use day to day: signals from neighboring muscles mixing together, drift and motion artifacts, electrodes shifting during or between sessions, muscle fatigue changing the signal, and models trained on short recordings that do not stay reliable. [6]

A wrong estimate is not a small error. If the controller thinks the wearer is in stance while the foot is actually swinging, it applies the wrong torque at the wrong moment. Good designs limit how much harm one wrong estimate can do.

Position, torque and impedance control

[1] [8]

The low-level loop and the gait-phase estimate may run on different processors, and every link between them, wired or wireless, adds delay. Timing matters because assistance is tied to a phase of a step that lasts a fraction of a second. In the exosuit study, the authors tested two timings for the start of push-off assistance and found that the best one varied between individuals. [3] [1]

The joint-level ideas behind these modes are explained in how robot joints and actuators work.

Worked example: a powered knee during stance

When the heel lands, the knee bends slightly as it takes the body’s weight, then straightens again as the body passes over the foot. A person with weak thigh muscles may find that early stance hard. A powered knee can add an extension torque during that part of the step and stay out of the way during swing.

[1] [3]

These four numbers are different, and the gaps between them are where problems hide. Gear friction, flexible cuffs and a knee axis that does not line up with the device all make delivered torque differ from commanded torque. A report that shows only the intended profile says nothing about what the wearer actually received. [9] [7]

Keeping the robot joint aligned with the human joint

A human knee is not a perfect hinge. Its rotation axis shifts as it bends, and soft tissue lets the cuffs move against the bone. Torricelli and colleagues found the largest mismatch between human and exoskeleton motion at the knee during gait, and point out that the human and exoskeleton knee axes differ. They list energy loss, less accurate control of the limb, and discomfort and pain from skin abrasion as consequences. [9]

Designers respond with adjustable link lengths, careful fitting and, in some devices, extra unpowered joints that let the frame follow the body. The test that matters is measuring interaction forces at the cuffs while the wearer walks, not only checking that the device’s own joints moved as commanded.

Safety: limits, stops and failure behaviour

[4] [2]

NIOSH lists muscle strain if a powered device moves beyond a joint’s normal range, burns from battery problems, pressure wounds and compressed nerves, and balance problems from a shifted centre of gravity among the potential risks of industrial exoskeletons. The US rule for medical exoskeletons requires verification of manual override controls where present and training in the use of safety features. [4] [2]

The safe state needs thought. Cutting all power to a shoulder-support suit is usually harmless. Cutting all power to a lower-limb exoskeleton that is holding a person upright can cause a fall. A stop in a wearable robot is a designed behaviour, not just an off switch.

How to read an exoskeleton result

The 2017 exosuit study is a good model of a careful claim. In nine people in the chronic phase of stroke recovery, powered assistance on the treadmill raised peak ankle dorsiflexion in swing by about 5.3° and reduced the energy cost of walking by 10%, compared with wearing the suit unpowered. Push-off on the weaker side rose by 11% in the seven participants whose push-off could be measured separately for each leg. The authors state that the study measured immediate effects only, not long-term therapy. [3]

• Who wore the device, and how many people took part?
• Was the comparison made against no device, an unpowered device or a different controller?
• Was walking measured on a treadmill, overground or in daily life?
• Does the report show delivered torque, or only the commanded profile?
• How were sensor failures, falls and stops handled during testing?
• Is the result an immediate effect or a change that lasted?