Why humanoid robots find hands and balance hard

TL;DR:

• A standing robot stays up while its center of mass projects inside the support polygon, the area under and between its feet.

• Reaching with an arm moves the center of mass, so the whole body must adjust while the hand is working.

• Hands are hard because they have many joints, depend on contact, and need force control as well as position control.

A person picks up a cup from a table while standing, without thinking about it. The same task for a humanoid robot involves two hard problems at once: do not fall over, and do not crush or drop the cup. Each is manageable alone. Doing both together is where humanoids earn their reputation for difficulty.

Balance: the center of mass over the feet

A robot standing still stays upright as long as its center of mass (CoM), projected straight down, falls inside the support polygon, which is the area covered by its feet and the space between them. If the projection leaves that area, gravity makes the robot tip over its foot edge unless it takes a step or applies some other force.

The CoM of a body made of parts is the mass-weighted average of the parts' positions: x_CoM = Σ(mᵢ xᵢ) / Σmᵢ. Moving any heavy part moves the CoM, which is why even a small arm motion matters.

Worked example: a humanoid picking up a cup while standing

Say the robot has a mass of 50 kg, stands with both feet flat, and each foot is 0.25 m long. Its CoM starts above the middle of the feet, so the margin to the front edge of the support polygon is 0.25 / 2 = 0.125 m. A cup of 0.3 kg sits 0.6 m in front of it. Track how the CoM moves. Forward is positive.

1. Arm moves out. The arm has a mass of 4 kg and its CoM moves forward by 0.3 m. The whole-body CoM shifts by 4 × 0.3 / 50 = 0.024 m.

2. Cup in the hand. The cup mass is 0.3 kg at 0.6 m, so the extra shift is 0.3 × 0.6 / 50.3 ≈ 0.0036 m.

3. Torso leans. To reach the cup, the robot leans its 25 kg upper body so that its CoM moves forward 0.10 m. The shift is 25 × 0.10 / 50.3 ≈ 0.050 m.

4. Total shift: 0.024 + 0.0036 + 0.050 ≈ 0.078 m, about 7.8 cm.

5. Compare with the margin: 0.125 − 0.078 = 0.047 m left, about 4.7 cm. The robot is still inside the support polygon.

That margin is thin. If the lean were 0.20 m instead, the torso term would be 25 × 0.20 / 50.3 ≈ 0.099 m, the total would be about 12.7 cm and the CoM would pass the toe edge at 12.5 cm. To reach farther safely, the robot has to move its hips back, step forward, or widen its stance. A controller that does this continually is estimating the CoM and the zero moment point (the point on the ground where the net tipping moment is zero), and adjusting the ankle and hip torques to keep it inside the foot.

Dynamics matter too. Accelerating the arm quickly pushes back on the body. Slow reaching keeps the static estimate honest. Fast motions need the full dynamic model.

Hands: many joints

A hand with five fingers can easily need 15 to 20 or more actuated joints. Each joint needs a motor or tendon, a sensor and wiring, all packed into a space the size of a fist. More joints mean more ways to move the fingers into a grasp. They also mean a bigger search for the controller, and more parts that can fail or wear.

Contact

Free motion in air is easy to model. Contact is not. When a fingertip touches a cup, the forces depend on friction, surface shape and how soft the finger is, and these change from one touch to the next. A small error of a few millimeters is the difference between a good grasp and knocking the cup over. Hands therefore need force and touch sensing, as described in how robotic hands sense touch.

The grip must also be strong enough. A 0.3 kg cup weighs 0.3 × 9.8 = 2.94 N. With two fingers and a friction coefficient of 0.5, each must press with at least 2.94 / (2 × 0.5) = 2.94 N. With a safety factor of two, aim for about 6 N. The hand has to hold that while the body keeps balancing.

Balance while reaching

The two problems interact. The arm's motion shifts the CoM, the balance controller moves the hips and ankles to compensate, and that moves the shoulder and so the hand. The hand controller then sees an unexpected motion at the cup. Humanoid controllers treat the whole body together, often assigning tasks in priority order: first stay balanced, then place the hand, then hold posture.

Working near humanoids: A full-size humanoid is heavy and can fall without warning. Test with a safety harness or overhead tether, keep people outside the fall zone, have a hardware emergency stop that someone is ready to press, and keep fingers away from joints and hands, which can pinch. Check battery packs for damage after any fall.

Where it goes wrong

• Slippery floors. Foot friction limits how much horizontal force the feet can apply.

• Unexpected payloads. A cup that is heavier than expected shifts the CoM more than planned.

• Sensing delay. If the balance loop is slow, the robot reacts after the problem has grown.

• Soft or uneven ground. The support polygon is not where the model says.

• Hardware limits. Joint torque and speed limits can stop a recovery movement from being fast enough.

How real robots use it

Many research humanoids are tested first in simulation, as described in why roboticists test in a simulator first, and on safety rigs before free standing. Some designs avoid the problem by using wheels, a wide base or a fixed stance. Wearable robots face the balance problem from the other side: the human supplies it, as in exoskeletons.

Try it yourself

Stand with your feet together beside a wall, with your fingertips near it for safety. Slowly reach one arm forward while holding a 1 kg bag. Notice how your hips shift back without you thinking. Then try on one foot. Your body is solving the same CoM problem with far better sensors than the robot has.

Frequently asked questions

Why not give robots wheels instead of legs?

Wheels are simpler, faster and more stable on flat ground, and many robots use them. Legs are useful where stairs, clutter or human-built spaces make wheels awkward.

Is the center of mass the same as the zero moment point?

Not exactly. In slow, steady poses their ground positions coincide. During fast movement they separate, and the zero moment point is the one that must stay inside the foot.