Why the motor leaves the knee

Every step swings the whole leg forward and stops it again. The torque the hip has to supply for that depends on the leg’s rotational inertia, and inertia grows with mass times the square of its distance from the hip. A motor halfway down the leg counts four times as much as the same motor a quarter of the way down. Mass low on the leg also adds to the unsprung mass that hits the ground at each footfall, which raises impact loads on gears and bearings.

The fix is old. A technology review of bipedal humanoid hardware by Ficht and Behnke traces it to early ASIMO, which used synchronous belts to couple off-axis joints and move actuator mass toward the limb root. ATRIAS, the spring-mass biped that preceded Cassie, put both leg motors at the hip and drove a four-bar linkage. Today the BHR8-J1 paper from Beijing notes that Unitree H1 and the electric Atlas place high-torque motors near the hip and reach the knee through linkages, and a 2025 paper by Lutz, de Matteis, Batto and Mansard describes a whole class of recent humanoids that follow Cassie in placing motors away from the joints to cut leg inertia.

Three ways to get torque down the leg

Once the motor sits at the hip, something has to carry its torque 30 to 40 cm down to the knee. Three families dominate, and many robots combine them with a conventional gearbox at the motor.

Timing belts are cheap, quiet and give a fixed ratio. USC’s HECTOR humanoid uses identical quasi-direct-drive actuators with a 9.1:1 planetary stage on every joint, and moves the motors toward the upper thigh and hip. Its knee is driven through a 1.55:1 timing-belt stage, so the total knee reduction is about 14.1:1. That turns 33.5 N·m and 21.0 rad/s at the actuator into roughly 51.9 N·m and 13.6 rad/s at the knee. The belt is both the remote link and an extra gear stage.

Cables and capstans wrap a rope or steel cable around drums of different diameter, which gives a reduction with no tooth contact and no backlash. ETH Zurich’s Capler-Leg drives both hip and knee this way with Dyneema DM20 rope: a ratio of 5 at the hip and 5.33 at the knee, including a 1.33 cable stage running from hip to knee. The 4.0 kg leg is rated at 70 N·m and 50 N·m/kg, and the authors report about half a million hopping cycles without visible cable damage. Capstans also appear in low-cost arms: the UMass Amherst Forte arm uses an 8:1 capstan at the shoulder yaw and a capstan-plus-belt stage of 18.75:1 at the shoulder pitch, for under $215 per arm.

Linkages push and pull rigid rods. A rotary motor near the hip turns a crank, and a pushrod moves a lever on the shank. This is the arrangement the BHR8-J1 paper attributes to Unitree H1 and the electric Atlas. The University of Bonn’s AGILOped goes further: two double four-bar parallelograms carry knee and passive-ankle motion from actuators at the hip, keeping the lower leg to 0.56 kg on a 14.5 kg robot. Linear actuators fit the same family. BHR8-J1 drives its knee with a ball-screw actuator and a guide-rod linkage tuned so the effective ratio changes over the stroke, peaking at a reported 318 N·m knee torque and supporting 0.5 m box jumps.

Published remote-transmission figures

How a belt stage works

A belt or cable stage multiplies torque by the ratio of the pulley diameters and divides speed by the same amount. In HECTOR’s case the knee pulley is 1.55 times the motor-side pulley, so for every 1.55 turns of the motor pulley the knee turns once. The belt teeth stop slip, and because the belt only works in tension, designers pre-tension it so both strands stay taut when torque reverses.

What a remote transmission costs

Moving mass up the leg is not free. Each family adds its own weak spot between the motor and the joint, and the controller has to live with it.

• Compliance. Belts and cables stretch under load. Capler-Leg’s Dyneema cable deflects about 2° at maximum torque, with a simulated joint stiffness of 1,614 N·m/rad. That is softer than a gear train and shows up as position error under load unless an output-side encoder corrects it.
• Nonlinear ratios. A linkage’s effective ratio changes with joint angle. Lutz and co-authors point out that controllers often approximate it with a constant ratio, which wastes part of the mechanism’s capability, and propose an exact but cheap analytical model for knee and ankle linkages. BHR8-J1 turns the nonlinearity into a feature, tuning it for high torque at take-off.
• Temperature and wear. Synthetic rope weakens with heat: the Capler-Leg team keeps DM20 well below 80 °C. Belts need tension that holds over thousands of hours, and linkages add pin joints and bearings, each a potential source of play.
• Coupling and packaging. A cable or rod that crosses an intermediate joint can tie two joint angles together, which the kinematic model has to account for. Linkages also limit range of motion, and belts need room for pulleys and tensioners in an already crowded thigh.

What it means for buyers and builders

• Read joint torque at the joint. When a belt or linkage sits between actuator and joint, the actuator’s catalogue torque is not the joint torque. HECTOR’s knee delivers 1.55 times its actuator’s torque at a lower speed.
• Ask where the encoders are. A motor-side encoder on a remote joint cannot see belt stretch or linkage play. Output-side sensing matters more when the transmission is long.
• Ask about service items. Belt tension, cable replacement intervals and pin-joint wear are maintenance items that a motor-at-the-joint design does not have.
• Expect more linkages, not fewer. With H1, electric Atlas and a wave of Cassie-inspired designs moving motors toward the hip, linkage modelling is becoming part of the standard control stack rather than a research topic.

FAQ

Sources