Falling is part of the job

Nobody building legged robots pretends otherwise. “Our robot falls a lot,” Boston Dynamics CTO Aaron Saunders told TechCrunch in 2024, “and one of the things we decided a long time ago [is] that we needed to build robots that can fall without breaking.” CEO Robert Playter put a number on it for Spot, the company’s quadruped: roughly one fall every 100–200 km of walking. Agility’s Digit ran at a 99% success rate over 20 hours of live demos in 2023 – and was still filmed going down.

So the useful question isn’t whether a humanoid falls. It’s what breaks when it does, and the answer usually starts inside the reducer.

The torque ladder inside a reducer

A gear catalog doesn’t give one torque limit. It gives a ladder. Rated torque is what the gear carries all day. Repeatable peak torque covers normal accelerating and braking. Above that sits momentary peak torque, which Harmonic Drive defines as the brief collision torque the gear may see “in the event of an emergency stop or collision”, to be kept rare. Go higher still and the failure modes change character: at 8 to 9 times rated torque, with the gear running, the flexspline and circular spline teeth can stop meshing properly and skip – ratcheting. At 16 to 17 times rated, with the input held still, the flexspline can deform plastically. That’s buckling, and the gear is scrap.

The pattern is consistent: a strain-wave gear shrugs off roughly four to seven times its rated torque, briefly, a limited number of times. Harmonic Drive even gives a formula for how many. The flexspline is allowed about 10,000 bending cycles under impact torque, so the permitted number of events is N = 1.0×10⁴ ÷ (2 × n/60 × t), where n is wave-generator speed in rpm and t is how long the impact lasts. A motor at 3,000 rpm taking a 0.15-second hit uses up 15 of those cycles per event, leaving room for about 670 events over the gear’s life. That’s a lot of emergency stops for an industrial arm. For a robot that falls every few days in a warehouse, it starts to look like a service interval.

Cycloidal reducers climb the same ladder with wider rungs. Nabtesco’s RV-E range allows acceleration and deceleration torque at about 2.5 times rated and is sized to take up to five times rated torque in an emergency stop – an RV-20E rated at 167 N·m tolerates roughly 835 N·m for an instant. That shock tolerance is one reason RV-type reducers get considered for the hips and knees of heavier humanoids, where a harmonic gear would have to be oversized to match.

Why gear ratio makes the hit worse

The foot stops in milliseconds. The motor rotor doesn’t. Its momentum has to go somewhere, and in a geared joint it goes through the teeth. The inertia a joint feels from its own motor is the rotor inertia multiplied by the square of the gear ratio, as Wensing, Kim and colleagues showed in their 2017 analysis of the MIT Cheetah actuator. Move from 10:1 to 100:1 and the reflected rotor inertia rises a hundredfold. Rated torque goes up tenfold, but the shock load from the motor’s own momentum grows much faster.

That’s why the Cheetah team settled on 5.8:1 and built a metric, the impact mitigation factor, to score how well a leg absorbs a hit. It’s also the case for quasi-direct-drive joints in humanoids: a low-ratio actuator gets pushed backwards by the impact instead of fighting it. High-ratio strain-wave joints, common in arms and in many hip designs at 50:1 to 160:1, can’t back-drive fast enough to dodge the spike, so they have to survive it. Motor maker CubeMars tells humanoid builders to size hip actuators for a 3–5× overload over rated torque for jumping and landing, and knees for 2–4×. That headroom is mass the robot carries every step.

A fuse in the gearbox

The alternative to carrying all that headroom is to let the joint give way on purpose. Both of the best-known American biped makers have patented the same basic idea: put a friction clutch inside the harmonic drive, preloaded by springs, so that above a set torque something slips instead of something breaking.

Boston Dynamics’ version, US12320400B2, granted in June 2025 with a priority date back in 2016, clamps the circular spline between two friction pads on the output member. A spring applies an axial preload, and in the patent’s words that preload “defines a torque limit”. Above it, the circular spline turns relative to the pads and burns the energy off as heat. The patent pairs this clutch-protected harmonic drive at the hip with a screw actuator at the knee, and it’s explicit about the motive: protection by slipping lets the designer use a smaller transmission than one sized to survive impacts by strength alone.

Agility’s US12416334B1, granted in September 2025 from a 2022 filing, does the same job one step earlier in the gear train. Clutch pads hold the flexspline still in normal running; Belleville spring washers set the pressure; above the limit, the flexspline slips. The drawings show a biped, and the text names shoulder, elbow, hip and ankle joints. Neither patent publishes a torque setting, and neither company has confirmed which production joints use it.

Other routes lead to the same place. Dutch start-up IMSystems pitches its Archimedes Drive, a rolling-traction reducer, partly on this point: under excessive force its rollers “can simply slide” rather than lock up or break, so a joint can be sized for nominal torque rather than for shocks. And industrial torque limiters already exist as separate parts – mayr’s EAS-compact, a backlash-free ball-detent clutch, starts at 5–50 N·m in its smallest size and re-engages automatically every 15°. What humanoid designers are doing differently is folding the fuse into the gear itself, where it costs almost no extra length or mass.

Software takes the first hit

Mechanical fuses are the last line. The first is the controller: current limits in the servo drive, and a policy that sees the fall coming and lands the robot on something other than its wrists and head. A 2025 paper, SafeFall, trained such a policy on a full-size Unitree G1. Compared with baseline behaviour, it cut peak contact forces by 68.3% and peak joint torques by 78.4%, and it avoided 99.3% of collisions with the parts the authors flagged as vulnerable, such as hands and sensors. A fall detector runs alongside the normal walking controller and only takes over once a fall is judged unavoidable.

Standards are heading the same way. ISO 25785-1, the safety standard for dynamically stable industrial mobile robots – legged robots included – is still in development at ISO. Controlled behaviour when balance is lost will be hard to leave out of it.

What to ask an actuator supplier

• The full torque ladder, not just rated and peak: momentary peak, and how many momentary events the reducer is rated for at your motor speed and impact duration.
• Reflected inertia at the output, in kg·m². It tells you more about fall survival than the gear ratio alone.
• Whether there’s a torque limiter in the module, where it sits, what the slip torque is and how much it drifts as the pads wear.
• Where position is measured. A motor-side encoder alone can’t see a clutch slip.
• What the drive does in a fall: current limits, reaction time and whether the joint can be commanded soft before impact.

FAQ

Sources