What a toe joint does in a walking robot

A toe joint is a pitch axis across the front of the foot, roughly where the human ball of the foot sits. It splits the sole into two pieces: a main plate carrying the ankle and heel, and a short toe segment in front. In one simulation study on Atlas, the toe link made up about a quarter of the foot length.

Without it, the only way to start the next step is to tip the whole rigid plate up on its front edge. The contact shrinks to a line. Most flat-footed bipeds avoid that moment altogether and walk with bent knees instead, shortening the stance leg so the foot can stay flat. That’s the familiar crouched humanoid gait, and it costs torque. Sven Behnke made the point in 2005 with his 74 cm robot Toni: bending the knee to shorten the stance leg “creates high joint torques”, so Toni stood on its forefoot and kept the stance leg straight. After long walks its knee servos ran coolest of all the leg joints.

The TU Munich team that designed LOLA gave the same reasons. Lifting the heel lets the swing leg stay more extended, and keeping the toe segment flat on the ground “stabilizes the robot and facilitates forward roll across the forefoot”, which should take load off the hip and knee. Their motivation was speed. Their earlier robot JOHNNIE topped out at 2.4 km/h, against 6 km/h for Honda’s ASIMO.

Active, passive or none

There are three ways to build the front of a humanoid foot, and each has a long research record.

An active toe puts a motor on the toe pitch axis, so the robot can push off rather than just roll. The University of Tokyo’s H6 and H7 were early examples; the LOLA designers noted in 2007 that they were the only humanoids with actively driven toes at the time. The newest is a 14-DoF biped from KOREATECH, KIST and WIRobotics, described in June 2026, which drives its 1-DoF toe through a coupled belt-and-linkage so the motor can sit higher up the leg. The catch is mass at the very end of the limb. A University of Tehran study on SURENA III found that a heavier toe link lowers the risk of tipping and slipping but costs more energy and joint torque.

A passive toe is just a hinge with a spring and a damper. It bends under load and springs back in swing. Waseda’s WABIAN-2 used one, tuned from human motion capture, to reproduce heel-contact, toe-off walking. The SURENA III team built theirs as a deliberately low-cost add-on and reported 15.3% less energy at the ankle and 9.0% less at the knee.

No toe means a single rigid plate. That’s still the commercial default. Unitree lists six DoF per leg on the G1, and Menlo’s Asimov 1 documentation says plainly that it “no longer uses passive toe joints” and now runs “fixed flat feet”, with a parallel ankle driven by two actuators.

What the numbers say

Hard numbers on toes are scarce, and few come from hardware. The cleanest comparison is the June 2026 paper by Kim and colleagues, which ran the same 14-DoF biped with and without its active toe at 1.33 m/s. With the toe, total power fell from 313.0 W to 260.2 W and cost of transport from 0.776 to 0.640 – 17.5% better. Heel-strike ground reaction force dropped 5.0%, and average path deviation in a steering test fell by a quarter. Not everything improved: the toe-less version hit a higher top speed, 3.55 m/s against 3.31 m/s. The authors stress that all of this awaits hardware validation.

For scale, the same paper puts a 70 kg human at a cost of transport of 0.316 at that speed, and says conventional bipedal robots usually sit well above 1. A toe joint closes part of the gap. Not most of it.

Older work points the same way with fewer figures. Agarwal and Popovic’s 2018 Atlas simulation showed a stronger thrust at toe-off and less knee bend in double support. The HRP-4C team at AIST used its toes to hold ground contact longer each stride so the legs could straighten. WABIAN-2RIII matched the human 12° of foot rotation and a 0.6 s step time, though only walking in place.

Why commercial humanoids still walk on plates

If toes save energy, the obvious question is why the robots being built in volume leave them out. No maker has published a full answer. Menlo’s documentation records the switch from passive toes to flat feet without giving the reason. But the trade-offs are visible from the engineering.

• Mass at the end of the leg is the most expensive mass on the robot. Every gram in the foot has to be swung forward on every step, which is why most new humanoids already pull the ankle motors up onto the calf through a parallel linkage. An active toe puts a motor, or at least a drive train, right back where designers worked to remove one.
• A toe joint carries close to full body weight at push-off, on a small hinge, in the part of the robot that hits the ground hardest. That makes it a durability and sealing problem as much as a kinematic one.
• Learned walking controllers are trained in simulation. A rigid plate is a simple contact to model. A sprung or actuated toe adds a contact mode and a stiffness value that both have to match the real hardware, or the policy misbehaves on transfer.
• The payoff depends on the job. Toes matter most for long strides and fast, efficient walking. A robot that shuffles between totes in a warehouse aisle spends much of its time standing, turning and taking short steps.

Tesla is the notable exception. When it showed Optimus Gen 2 in December 2023 it credited “torque sensing, human foot geometry, and articulated toe sections” for a 30% faster walk, without saying whether those toes are driven or sprung.

What to ask a vendor

Ask how the foot is built before you ask about walking speed. Is there a toe joint, and is it driven or sprung? If it’s sprung, can the stiffness be changed, and is it in the simulation model the vendor ships? If there’s no toe, ask for the robot’s walking energy figure – watts or cost of transport at a stated speed – because a flat-footed robot pays for it at the knees and hips. And for any foot with moving parts, ask about the service interval. The toe is the joint that sees the most dirt, water and impact on the whole machine.

FAQ

Sources