Gearbox-FreeArms Built forRobot Learning
Transverse-flux direct drive deletes the gearbox from every joint — demonstrations at full human speed, an actuator your simulator can actually model, fleets that don't drift with wear.
TORQUE DENSITY TARGET · REMOVES THE GEARBOX
≈6–10×
WORKING PROTOTYPE
3 N·m
TORQUE500
RPM⌀60
MMDIRECT DRIVE — NO GEARBOXCLEAN DATAENABLES ZERO-SHOT SIM-TO-REALNO WEAR DRIFT
01 / THE GEARBOX IN THE LOOP
Your model learns the gearbox, too
Every demonstration you record and every policy you transfer passes through a gear train first — backlash, stiction, a speed cap, and wear, written into every sample. Your model fits all of it, faithfully.
ONE COMMAND, FOUR ARTIFACTS · MECHANISM SCHEMATIC, EXAGGERATED
02 / ONE COMPONENT, FOUR ARTIFACTS
One component, four artifacts
The gearbox trades away speed and corrupts position, force, and consistency — four artifacts per joint, multiplied by every joint in the arm. And the map from commanded current to delivered torque becomes a function nobody can write down exactly, one that changes as the teeth wear.
GEARBOX
S01
Speed
Reduction trades speed for torque
S02
Position
Backlash at every reversal
S03
Force
Nonlinear friction — current ≠ torque
S04
Consistency
Wear drift — different on every unit
03 / THE WORKAROUNDS
You're already paying for the gearbox
Harmonic gears, real-world RL, or learning it as part of the embodiment — three workarounds, three different bills. Building the body in-house doesn't escape them — those joints still need one of the three — it just adds the biggest bill: time and resources that could have gone into data and models.
OPTION A
Harmonic drive + torque sensor
Buy precision backYOU PAY IN
Price + agility
OPTION B
Real-world reinforcement learning
Train around itYOU PAY IN
Dataset size + iteration speed
OPTION C
Learn it as part of the embodiment
Model it outYOU PAY IN
Dexterity
THE FOURTH WORKAROUND — BUILD IT YOURSELF
· Tesla· OpenAI· Figure· Boston Dynamics· Physical Intelligence· 1X· Dyna· Genesis
We keep the body state-of-the-art — you keep your team focused on data and models, for max speed and performance.
04 / DELETE THE GEARBOX
Remove the gear train and the artifacts go with it
Vassar transverse-flux motors put torque directly at the joint: full speed, zero backlash, current that reads as torque, nothing meshing to wear. The artifacts aren't compensated, calibrated, or learned around — they're gone.
DIRECT DRIVE · NO GEARBOX
S01
Speed
No reduction → full speed
S02
Position
No teeth → zero backlash
S03
Force
No train → current ≈ torque
S04
Consistency
No wear → no drift
No gearbox in the joint. No gearbox in your data.
05 / TRANSVERSE FLUX
Removing the gearbox was never the hard part — torque density was
A gearbox-free joint only works if the bare motor makes working torque: radial flux tops out near 0.3–0.5 N·m in the humanoid joint envelope — our design targets ~3 N·m, on the order of 6–10×, and the first bench prototype already delivers it: 3 N·m rated, 500 RPM, ⌀60 mm. You're going to verify our numbers anyway — we'd rather you start from accurate ones.
RADIAL FLUX
0.3–0.5
VASSAR TRANSVERSE FLUX
3.0
0
1
2
3
≈6–0×
TORQUE MULTIPLIER — TARGET06 / WRIST TO HIP
One architecture, from wrist to hip
Off-the-shelf actuators climb a steep diameter-versus-torque curve — more torque means a bigger motor or a bolted-on gearbox, and a different dynamics model at every joint. One family, sized to the joint, means one actuator model, one control interface, one set of assumptions from wrist to hip.
Their torque curve needs a gearbox. Ours is the bare motor.
07 / CLEAN DATA
The joint doesn't edit your demonstrations
With no teeth in the path, demonstrations run as fast as the operator can move — dexterous, contact-rich, and the demo you give is the trajectory the robot records. And because current maps to torque, a usable force signal comes free with every demonstration.
08 / ZERO-SHOT SIM-TO-REAL
Physics clean enough to simulate faithfully
No backlash, no stick-slip, no load-dependent mesh friction — what's left is low-order motor physics your simulator can model in high fidelity. The closer the simulated joint sits to the real one, the smaller the gap your policy has to cross: the actuator side of zero-shot transfer, addressed at the source. Domain randomization goes back to covering the world, not your own transmission.
THE ACTUATOR MODEL, SIM vs REAL
WHAT A GEARED JOINT DELIVERS
τ = f(u, θ, θ̇, load, wear, …)
Backlash, stick-slip, load-dependent mesh friction, aging — hysteretic physics that resists faithful modeling.
WHAT A GEARBOX-FREE JOINT DELIVERS
τ ≈ k·i
One motor constant, to first order — small, smooth corrections you can identify on a bench.
WHAT THAT BUYS YOU
sim ≈ real, at the joint
A joint your simulator can model in high fidelity — the actuator side of zero-shot transfer, addressed at the source.
09 / RELIABILITY
The fleet you deploy stays the fleet you trained
Gear teeth wear differently on every unit, so a geared fleet slowly walks away from its own baseline. With nothing meshing to wear, month 12 behaves like day 1 — the plant doesn't wear out from under your policy.
10 / HARDWARE + SOFTWARE
We do the plumbing — you keep the models
Firmware, drivers, a torque-level control API, sim-ready models — the layer between your training stack and the copper ships with the hardware. Your data and models stay yours: we're a component in your system, not a platform you have to join.
YOUR MODELS & DATAYOURS — STAYS YOURS
Policies, datasets, training stack
Portable and under your control. We never see your data unless you hand it to us.
VASSAR SOFTWARESHIPS WITH THE HARDWARE
Control API · sim-ready models · tools
Torque-level control, simulation models that match the physical joint, calibration and logging out of the box.
VASSAR HARDWARETRANSVERSE FLUX · NO GEARBOX
Gearbox-free joints & arms
Full speed, zero backlash, current ≈ torque — the signal layer your stack sits on.
We build the joint. You build the intelligence.
11 / WHAT WE'RE BUILDING
Arms first — then the full body
A gearbox-free arm ships first — torque-level control, simulation-ready, software included. Then the same gearbox-free joints scale into the full body, and the actuator models, tooling, and control interface you build on the arm carry over. If you want your hardest task on early hardware, or to shape the spec while it's still movable, talk to us now.
PHASE 01 · NOW
Gearbox-free arms
PHASE 02
The full body
- WHAT YOU GET
- Gearbox-free arms + the full software stack
- BUILT FOR
- Imitation learning, RL, and deployment
- WHERE IT STANDS
- Working prototype joint — 3 N·m · 500 RPM · ⌀60 mm
- WHAT CARRIES OVER
- Actuator models, tooling, and integration transfer to the body
12 / WHY VASSAR
Built by people who have shipped robots before
200+ robots delivered, patents on an implementation of transverse flux that scales, US R&D with DFM and manufacturing in Asia. Designed to be made, not just demoed — when your stack depends on a joint, the supplier's ability to keep making it matters.
200+
Robots delivered
Patented
Transverse flux that scales
US + Asia
R&D and manufacturing
BUILD STACK· Al· Fe· Cu· Magnets· CNC· Wire EDM
13 / GET IN TOUCH
Tell us what you're training
Write to us with what you're building and what your joints need to do — we'll tell you exactly where the hardware stands, measured numbers and honest targets included. Real technical questions get real technical answers.
LABS
Frontier robotics and embodied-AI labs — pilot the arms, shape the spec.
DEVELOPERS
Individual builders who fight gearbox artifacts daily — get on the early list.
TALENT
Engineers who want to build the joints, arms, and body with us.