A Long-Legged, Direct-Drive Monopedal Robot Achieves Exceptional Jump Height

TL;DR: Every jumping-robot record until now has been set by machines that slowly wind up a spring and release it with a latch — because a small electric motor simply can't deliver enough instantaneous power through a rigid drivetrain. Na and Yim show that with the right transmission you don't need the spring. Their 281 g single-legged robot jumps 7.6 m straight up with a 190 ms ground-contact time — the highest jump ever for an electrically actuated, spring-free robot, and the highest "vertical jumping agility" (height × jump rate) of any electrically actuated jumper, springs included. The enabling idea is a fabric tape that winds onto a motor spool: as fabric accumulates, the spool's radius grows, so the gear ratio continuously shifts through the jump — a 16th-century clockmaker's trick (the fusée) reborn as a self-shifting CVT that holds a tiny motor at peak power for the entire leg stroke.

Why springs have ruled jumping

To jump high, you need to inject a lot of kinetic energy in a very short leg stroke. A motor's energy is essentially unlimited (it comes from the battery), but its power — energy per second — is capped. The classic dodge is elastic energy storage: use a huge gear reduction to slowly load a spring over seconds or minutes, hold it with a latch, then release everything at once. Hawkes et al.'s famous 32 m jumper works exactly this way, with a 1000:1 reduction — and pays for it with roughly 120 seconds of winding between jumps, plus latch and trigger mechanisms, and essentially no ability to modulate force mid-jump.

The alternative — direct drive (DD), where the motor is rigidly coupled to the leg — is beautifully simple and controllable: ground force tracks motor torque instantly, you can command force at high bandwidth, and you can even estimate leg force from motor current alone (free proprioception). But the jump can never exceed the motor's instantaneous power. For a power-limited actuator without air drag, jump height scales as

where l_{\text{leg}} is the leg stroke, P_{\max} the motor's peak power, and M the total mass. This equation is the whole design brief of the paper: if you refuse springs, your only levers are a longer stroke and a better power-to-weight ratio. So the authors build a robot that is almost nothing but leg — a 1981 mm carbon-fiber tube weighing 27 g, giving an enormous 1867 mm stroke — driven by a 53 g brushless motor.

Two problems immediately appear. First, a 2 m tube with 6 mm outer diameter and 0.5 mm walls will buckle like a dry spaghetti noodle under jump loads (a back-of-envelope from their numbers: accelerating the 192 g body to roughly 18 m/s in 190 ms means ~18 N of axial force, about 10 g's on the body). Second, a fixed-ratio transmission wastes most of the motor. A DC motor produces peak power in a narrow speed band; with a fixed gear ratio, motor speed simply tracks body speed, sweeping from zero through the entire speed range in one stroke. The motor spends the crucial early stance producing high torque at near-zero speed — that is, near-zero power.

The fabric-wrap transmission solves both problems with one piece of tape.

A fusée made of Dyneema tape

Old spring-driven clocks faced the mirror image of this problem: a mainspring's torque droops as it unwinds. Clockmakers compensated with the fusée — a cone-shaped spool that changes the effective moment arm as the cord winds, keeping delivered torque constant. Na and Yim get the same effect without machining a cone: the spool's radius grows automatically as the fabric winds onto itself.

Concretely: a Dyneema-composite tape, 6 mm wide and 0.13 mm thick, is anchored near the top of the leg and spirals down the tube at a shallow angle \phi = 2.5° from the leg axis. Its other end winds onto a spool on the motor shaft inside the body, which sits at the bottom of the leg. Spinning the spool pulls the body up along the leg — that's the jump stroke. Since each wound layer of fabric adds thickness, the spool radius grows with leg extension x:

where r_0 is the initial spool radius and t the fabric thickness. The mechanical advantage — motor angular velocity per unit body velocity — is \mathrm{MA}(x) = 1/(r(x)\cos\phi). Plugging in their parameters (r_0 = 2 mm, t = 0.13 mm, 1.867 m stroke), the spool grows to about 9 mm by liftoff, so the effective gear ratio drops roughly 4.5× over the stroke. High ratio at the start, when the body is slow and needs force; low ratio at the end, when the body is screaming along and the motor would otherwise redline.

And the same tape, spiraled tightly around the tube, does double duty: it braces the absurdly slender leg against buckling and damps its first bending mode. The transmission is the exoskeleton. The authors chose the geometry by simulating stance dynamics with a current-capped DC motor model and aerodynamic drag, sweeping r_0, t, and \phi; the optimum wanted an even smaller r_0 (1 mm) than they built, but 2 mm was chosen for manufacturability at negligible predicted height loss.

Pinning the motor at its power peak

Here's the operating-point picture that makes the design click. A current-limited DC motor has two regimes: below a "knee" speed \dot\theta_{\text{knee}} = K_v(V_s - I_{\max}R), current saturates at I_{\max}, torque is flat, and power rises linearly with speed; above the knee, back-EMF chokes the current and power falls off toward the no-load speed.

With a rack-and-pinion (as in the authors' earlier robot, which managed 0.59 m), the motor crawls through the low-power start of its curve while the body accelerates. With the fabric wrap, the initial 500 rad-per-meter ratio lets the motor spin up toward its knee almost immediately, and the automatically declining ratio keeps it parked there. The measured motor telemetry (Fig. 8 of the paper) confirms it: after an initial ramp, mechanical output power P = K_t I \dot\theta sits near its maximum for the rest of the 190 ms stroke. That is the entire trick — the same motor, wrung out for nearly its full peak power over nearly the full stroke, no gearbox, no spring.

One subtlety worth savoring: because the no-slip wrap makes the whole system a single degree of freedom, the dynamics reduce cleanly to \ddot x = (\tau\,\mathrm{MA} - J\,\mathrm{MA}\,\mathrm{MA}'\dot x^2 - mg - F_d)/M_{\text{eff}}(x) with an effective mass M_{\text{eff}}(x) \approx m + J\,\mathrm{MA}^2 — the reflected rotor inertia J\,\mathrm{MA}^2 is enormous at the start (when MA is high) and shrinks through the stroke, which is exactly the CVT-like behavior you'd design by hand.

The leg as a lever for tiny propellers

A 2 m monopod is an inverted pendulum from a nightmare, so how does it stand up? By exploiting the leg's length a second time. A 55 g balancing module at the top of the leg carries four 7 mm coreless motors with 76 mm props, an IMU, and its own microcontroller. Total thrust authority is a mere 0.50 N — 18% of the robot's weight, useless for flying — but acting through a ~1.97 m moment arm about the foot, it's ample torque for attitude control. A computed-torque PD law cancels gravity during stance (F = -\frac{1}{L}(k_g\sin\psi + K_p\psi_e + K_d\dot\psi_e), where \psi is lean from vertical and L the foot-to-module length); at liftoff the controller switches gains because the rotation axis jumps from the foot to the center of mass.

This buys real capabilities: the robot balances upright before a jump, tracks commanded leans up to 15° (tilted jumps at ~20° lean were demonstrated, steering the takeoff direction), and reduces attitude deviations in flight. But aerial control is weak by construction: after liftoff the body sits at the top of the leg next to the thrusters, so the moment arm L - x_{\text{COM}} collapses. Repositioning the body along the leg mid-flight — which would restore the arm and also re-cock the leg for the next hop — is explicitly not yet implemented.

The evidence

The headline jump was measured outdoors by stereo triangulation from two calibrated 120 fps cameras with meter-stick references: peak body height 7.6 m, flight lasting ~2.5 s, horizontal drift under 1.6 m. Stance time was ~190 ms. On the paper's own agility metric — jump height divided by (stance + ascent time), readable as "how fast could this thing climb stairs sized to its jump" — it scores 5.30 m/s.

The two charts together tell the real story. Hawkes et al. still owns raw height by 4×, but at 0.25 m/s agility — it's a mortar, not a leg. This robot is the first spring-free machine in elastic-jumper height territory, and it tops the agility chart outright, beating even the best spring-based hopper (Bai et al., 3.93 m/s, and that number includes quadcopter assistance). The previous best direct-drive jump heights were 1.6 m (RAMIEL, at 10.3 kg) and 1.14 m (Stanford Doggo) — this is a ~5× jump-height improvement for the actuation class, at a fraction of the mass.

What to make of it, and what to be skeptical about

The result is a clean existence proof for a claim many would have doubted: you can get elastic-jumper-class height from a bare motor and battery if the transmission does its job. Since direct drive retains everything springs give up — continuous force control through stance, high-bandwidth GRF commands, per-jump energy modulation, current-based proprioception, and near-zero recovery time — this reopens the design space for agile hopping machines. The scaling law says the recipe (long stroke + high P/M + a transmission that actually delivers P_{\max}) should transfer; the authors plausibly suggest self-launching fixed-wing drones and low-gravity planetary hopping, where a 7.6 m Earth jump translates into crater-rim-clearing leaps on the Moon.

Now the caveats, which are real. The paper validates the flagship jump as a single hardware experiment — no repeatability statistics, no distribution of jump heights, no cycle-life data on the fabric (a 0.13 mm tape taking ~18 N loads and abrasive wraps every jump is a plausible wear item). The robot does not hop: after landing there is no demonstrated way to re-cock — the body must travel 1.87 m back down the leg, and mid-air repositioning is future work, so the "agility" number is currently a projection of what continuous hopping would achieve, computed from one jump's stance and flight times, not a demonstrated locomotion rate. Landing survival is asserted ("sufficient durability to survive jumps and landings outdoors") but not characterized — a 281 g machine falling 7.6 m, even leg-first, is a nontrivial impact. Aerial attitude control only "reduced deviations but did not fully settle them." And the height is measured at the body, not a true CoM, though their argument that 247 of 281 g are co-located at apex makes the correction genuinely negligible.

None of this undercuts the core contribution. The fabric-wrap transmission is one of those mechanisms that seems obvious only in retrospect — a variable-ratio drive with zero extra parts, whose "gearbox" is the wound tape itself and whose structural bracing comes free. Section II-B is the part of the paper to read carefully: the derivation from wound-fabric geometry to r(x), MA(x), the effective-mass dynamics, and the design optimization is short, elegant, and complete enough to reimplement. If you only look at one figure, make it the paper's Fig. 8 — flat peak power across the stroke is the whole thesis in one telemetry trace.