
A really underrated part of actuator design is where the connectors sit on the driver board.
A lot of driver boards put them on opposite sides, in on one side and out the other. But in a humanoid joint with big ROM, the link sweeps right over one of those cables, which makes a clean hidden harness very annoying.
Putting both connectors next to each other lets you angle the input and output toward the part the joint never sweeps through, so you need a lot less wasted space to route and protect the harness.
I'm curious why opposite sides became the default for a lot of actuator companies.
Been waiting on the Unitree H2-X specs, and wow. The base model sells for $25k and likely costs Unitree ~$9k to build (their IPO prospectus puts humanoid gross margin at ~63%, could be lower here).
Fully loaded with a Jetson Thor ($3k list) and their own dexterous hands, a 31 DoF humanoid probably costs Unitree $15 to 20k. That's below Tesla's long term retail target for Optimus ($20 to 30k). Unitree could sell at Tesla's dream price today and still make money on the hardware.
The specs are insane as well. Dual hot swappable batteries (Figure and Optimus each run a single built in pack and have to dock to charge), arm payload above Figure/Optimus (~30 kg peak across both arms, vs ~20 kg listed for each). All at roughly a third to a quarter of the cost of what Figure and Tesla are building, by most estimates.
We're legit at the point where Unitree could dominate the global humanoid market if they crack intelligence, or even just build a teleop stack good enough to arbitrage labor. With cheaper operators, Chinese robots can do real work today with humans in the loop, which likely gets the data flywheel spinning faster than we can match, especially now that new Chinese humanoids are banned here.
And the only practical way to even prototype a similar design in the US is importing almost everything from China:
No low MOQ harness makers here for multi stranded through bore wiring. No hollow shaft motor controller boards, let alone ones with connectors oriented for through bore routing and hidden wiring. No US integrated actuator anywhere near humanoid torque at a reasonable price.
Structural prototyping costs ~5x more here even after tariffs, and US PCB fabs run several times the price with longer lead times. Even batteries: there are no low MOQ US shops doing custom hot swap packs with a BMS anywhere near Chinese prices.
Same with data collection hardware: hand exo FPCs, cheap camera boards, small custom encoders. All mass produced in China for a fraction of the price, and nearly impossible to prototype here.
It’s why so many US humanoid startups are just reskinned Unitree G1s. We gotta lock the hell in.
1
2
220
We’re WiscoHumanoids, a student team at UW Madison.
Our goal is to have the humanoid fully built and running by the end of the school year. Funding is one of our biggest constraints right now, so we're looking for sponsors to help us move faster.
And if you’re hiring, our team has some of the best builders I know. DMs open
Big thanks to our current sponsors @UWMechEngr @UWMadisonECE @UWMadEngr @sendcutsend
Our humanoid is 31 DOF: 14 in the arms, 12 in the legs, 3 in the torso, and 2 in the head. We’re building the arms and legs now, with significant work going into ROM, kinematics, and hardware co-design. The structure uses a mix of bent sheet metal, flat sheet, and billet rather than machining everything from billet like Unitree, to minimize costs.
On electronics, for rev 1 individual mechanism prototyping, we’ve designed a custom PDB that takes 38 to 50 V input and powers the actuators, IMU, and Orin, with precharge, regen protection (2.82 mF bulk capacitance plus braking resistors), and a hardware e stop contactor. During testing it’s fed from an off the shelf PSU. Each limb’s motors run on their own CAN FD bus through USB to CAN adapters into the Orin, so we can bring up each mechanism independently. In parallel, we’re working on the electronics for the full robot. We’ve laid out an EtherCAT version with custom junction boards, and we’re evaluating it against scaling up our CAN FD setup before committing, since EtherCAT needs a junction for each Y split. The compute stack is built around an RK3588S board paired with a Jetson Thor and a quad GMSL capture card for cameras. We’ve specced custom harnesses that route through the joint bores and are ordering them built to our drawings. We’ve also designed a hot swappable 12S2P battery pack using Molicel P50B cells; its mechanical design is done, and we’re now working on the BMS and FPCs.
On the learning side, we’ve trained a gearbox assembly policy on our Panthera arms, with insertion reliability still improving under tight tolerances (0.1 mm, 4 dowels). We’ve also implemented SuperMimic and other RL methods in sim, alongside actuator characterization for sim to real.
Along the way, we’ve built on open source designs like YUBI, ALOHA Mini, and Aero Hand to learn from and improve our own hardware. Happy to answer any questions!
Hmm. Quite a bit of changes since H2 actually.
Knee is now directly actuated at the joint again compared to H2 (Linkage Driven). Guessing the linkage constrained deep flexion + more failure points.
Near neutral hip abduction actuator(J1) vs a high anhedral angle for the old H2. They’ve managed a much more compact hip box than G1, potentially improving inward adduction clearance for crossover stepping without needing an anhedral angle.
Ankle looks to have returned to G1's dual pushrod parallel linkage arrangement, rather than H2's cables for roll / push rod for pitch arrangement. Shin packaging is noticeably more compact, possibly enabled by longer linkages while preserving ankle ROM. Parallel linkage setup also brings back torque sharing for ankle pitch.
Dual hot swappable batteries (back face rather than side face extraction)
Surprising how much the leg architecture is still changing within Unitree’s designs vs the Arm still being serial. Would’ve expected more convergence by now, especially with kinematic codesign (Though I'm sure other factors play a role too).
What if one of the biggest bottlenecks in robot learning isn’t the policy, but the robot arm itself?
In sim, OpenArm misses 64% of the UMI episodes. YAM 42%. Axol 29%.
I searched 5,040 tabletop arm designs (red in the video = a miss). The best one misses just 9%, with 9 to 21× lower tracking error and 6 to 11× lower joint-speed penalties than those three arms.
And even when the other arms can replay a demo, they still pay for it: 8 to 15× higher smoothness penalties and 7 to 33× higher singularity-conditioning penalties.
Caveats: sim only, same controller for every arm, no learned policies tested yet. Full methodology in the article.
Makes me wonder how much arm kinematics biases which behaviors policies can actually learn from human demonstrations.
Anyone tryna talk UMI grippers? Been thinking through the optimal architecture.
Generalist's gear based gripper allows for the smallest form factor, but locks you into small servos, and removes the parallelism in gripping (Unsure how important parallelism really is? With the compliant finger tips they use, it absorbs the error)
One linear rail, dual carriage gear based/linkage based gripper allows for very thin height, but sacrifices width. (Worst design IMO? The extra rail length makes it easy to catch on objects)
Two linear rail, dual carriage gear based/linkage based grippers allows for low width, but sacrifices height.
Trigger based designs sacrifice fine grain movement, human affordances, and form factor sizing.
With the parallel jaw grippers its simpler to make it work for production with one 10Nm actuator vs gear based grippers generally use servos to not sacrifice form factor (which has torque issues depending on the task and cycle life issues depending on servo)
However, with the parallel jaw grippers, its much harder to make a non trigger based design with the same width profile as Generalist, as the extra height with the dual linear rail sacrifices ergonomics.
To me it seems like Generalist and 2 Linear rail dual carriage designs is the best, but the 2 Linear Rail design requires a non straightforward, heavier UMI design.
Ngl, this could be the endgame archetype for data gloves
Beats cameras on occlusion, IMUs on drift, EMF on interference, and exoskeletons on mechanical complexity
Mean fingertip error is ~4.9 mm, but most of that seems solvable: calibration, glove fit, and hand model error
This is insane value now that the J288 servo is $89. FOC brushless, dual encoders, Impedance control, Torque telemetry
Its closest Dynamixel (XC430, $138): brushed, single encoder, position control only
Servos are finally getting an architectural upgrade
Microduck is a really fun robot. We highly recommend giving our brushless digital servo a try. 🥰🥰unitree.com/DigitalServo
Readers added context they thought people might want to know
Microduck is not using Unitree's Servo.
x.com/matth_lapeyre/…
2
6
1,074
Both robots’ legs converged on essentially the same architecture: hip rotation → abduction → flexion, two actuators stacked at the hip flexion axis, one driving the hip and one driving the knee through a linkage, and the ankle driven by linkage from actuators at the knee.
All of it pulls actuator mass proximally, reducing distal leg inertia during swing.
The gap probably doesn’t come from the overall leg topology. It’s mass distribution, actuator design, linkage geometry, the arms, limb lengths, and the policy.
1
52











