I want to talk about the experience of our team Robonine in developing a manipulator with 6 degrees of freedom. The project is currently at the prototype testing stage, but during this time enough material has accumulated to share with the community — from the motivation and the choice of kinematic scheme to fighting backlash in budget servo drives and topological optimization of the structure.
Manipulator prototype without the cover
How it all started
In early 2025 I was strongly hooked by the Aloha Mobile project video by researchers from Stanford and Google. In it their mobile platform autonomously watered flowers and washed dishes. Even a year ago this looked like substantial progress in robotics — imitation learning on real tasks in a home environment.
The project is open, I went to look at the components — and the price greatly surprised me.
Cost of existing solutions:
- Aloha 2 (tabletop assembly of 4 manipulators, Trossen Robotics): ~$19,000
- Aloha Mobile (mobile platform): ~$26,000
- Agilex Cobot Magic: ~$30,000
Aloha Mobile — mobile platform from Trossen Robotics with 4 manipulators
There are other similar developments, but the price range is approximately the same. For a research lab this might be acceptable, but for broader application — the barrier is too high. We decided to try to make it significantly cheaper.
Choice of actuators: why servo drives
The key and most expensive component of a manipulator is the actuators. This is approximately 60% of the cost of components (BOM). Accordingly, if the goal is to make a cheap manipulator, you have to start precisely with them.
Of the options in the budget segment:
- Stepper motors — good torque, but no position feedback without an additional encoder, prone to skipping steps under load.
- Servo drives (smart servo) — built-in encoder, feedback, possibility of serial connection (daisy chain).
The price range is approximately the same; in different budget projects both are used. We settled on servo drives.
Feetech STS3215
Feetech STS3215 is one of the most popular budget smart servo drives. For $15–30 per unit we get:
- Magnetic encoder at 12 bit (4096 positions)
- Working torque around 15 kg·cm
- Weight 55 grams
- Serial connection over a bus (TTL/RS485)
Among the DIY community the Feetech STS3215 model is especially popular; it is used in such open projects as the 5-axis SO-ARM100. We also chose this line as the basis.
Choice of kinematic scheme: semi-SCARA
The next important decision is the choice of kinematics. A classical anthropomorphic manipulator with a long arm (say, 650 mm) for a load of 1 kg creates a torque:
M = F × L = 1 kg × 9.81 m/s² × 0.65 m ≈ 6.38 N·m ≈ 65 kg·cm
on the servo drives at the base. With budget servos with a torque of 15 kg·cm this is simply an unworkable scheme. To solve this problem, we chose a semi-SCARA design:
- Vertical movement — a linear unit with a ball-screw transmission (by analogy with 3D printers).
- The SCARA part — the links work in the horizontal plane, which removes the load from the weight of the structure off the drives of these axes.
Lego prototype to validate the kinematic concept
We started with Lego Technic — a fast way to test the kinematics idea.
CAD model of the first prototype — view of the semi-SCARA design
CAD model of the first prototype: visible is the tower with a ball-screw transmission providing motion along the Z axis, then there are two cantilevers that move only in the horizontal plane (the SCARA part), and the manipulator forearm with the gripper.
The problem of backlash in budget servo drives
This is the moment we initially underestimated. When we started testing the first prototype, it became clear that backlash in budget servo drives is a serious problem.
Backlash — the gap between gear teeth when the direction of rotation changes
Our measurements showed that for budget-category servo drives backlash can differ substantially from the datasheet values: 0.8° instead of the declared <0.5°. It would seem to be a trifle — but on a 650 mm arm an angular error of 0.8° at one joint gives a linear deviation of the order of 9 mm. And if you sum the backlashes of all joints, add structural compliance, backlashes in coupling units, and the limited precision of the encoder — the absolute positioning error becomes already greater than 1 cm. Such an error is very difficult to compensate for in software.
Backlash compensation by paired servo drives
To solve this problem we use a combination of two servo drives on each axis. The motors operate with a small preload relative to each other — one "pushes" clockwise, the other — counterclockwise. This practically completely eliminates the dead zone (backlash).
CAD model of the unit with two servo drives for backlash compensation
The diagram above shows the design of a unit with paired servo drives; backlash compensation is performed at the expense of the preload created by the second motor. It is important to note that this approach works well with a small load, since the increased friction of the gears noticeably accelerates wear of the gearbox. Nevertheless, for our task (the manipulator is not industrial, the loads are moderate) — this is a good compromise.
Bench for measuring backlash
Measuring backlash correctly is a separate task. The problem is that backlash is very easy to "take up" when you touch the gear with a measuring instrument. Therefore we assembled a special setup which uses a rubber band that provides a minimal preload without distorting the results.
Bench for measuring backlash: a rubber band provides the minimal preload that does not distort the result
A detailed description of the measurement methodology and results is described in an article which is currently going through the final stage of review at the HardwareX journal.
Second version: dense layout
In the image below — one of the most complex units of the second version of the manipulator. Here the servo drives are arranged in pairs to work with two perpendicular axes.
Mock-up of the unit assembled from acrylic glass
Dense layout of the second-version unit: pairs of servos work on two perpendicular axes.
Layout of all units
The full design has 11 servo drives and 1 stepper motor (for the vertical lift with a ball-screw transmission).
Layout of all 11 servo drives and the stepper motor
The photo above shows the numbering of the manipulator's drives: 11 servo drives + 1 stepper for the linear axis.
Test assembly
Unit made of sheet metal. Video in motion
Next our engineer moved on to testing the operation of the unit under load on an improvised bench.
Cost optimization: from 46 parts to 6
The assembly was sent to manufacturing for evaluation. Custom metal parts alone numbered 46 items (laser cutting, CNC aluminum machining, lathe work). The total cost for single-piece production was on the order of 150,000 rubles (~$1,500). The retail price came out close to what Agilex offers with its Piper 6 DOF — that is, the price advantage was lost.
Such a cost did not suit us. At the next stage we radically simplified the design. We considered several options:
Manipulator design from aluminum profiles
Radical simplification along the lines of Aharobot (minimum parts, maximum compromises on stiffness).
Stiffer variant of the 6DOF manipulator
Above is shown the stiffer variant preserving the key design decisions, on which we decided to settle.
Cost together with shipping from China amounted to $452
In the end the number of custom parts was reduced by a factor of 7 — from 46 to about 6 items. Cost together with shipping from China amounted to $452. Total BOM cost — about $900.
Topological optimization
Topological optimization of structural stiffness. Full article
After simplifying the design we carried out topological optimization of the key load-bearing elements. Results:
| Parameter | Original design | After optimization |
|---|---|---|
| Mass | 1.937 kg | 2.376 kg (+22%) |
| Max. stress | 93 MPa | 25 MPa (↓ by 3.7×) |
| Deviation along the Y axis | 1.05 mm | 0.41 mm (↓ by 2.5×) |
| Deviation along the X axis | 1.03 mm | 0.31 mm (↓ by 3.3×) |
| Deviation along the Z axis | 0.62 mm | 0.21 mm (↓ by 3.0×) |
With a moderate increase of mass by 22% we got a reduction of the maximum stress by almost 4 times and a reduction of deviations along the axes by 2.5–3.3 times. Detailed material on the optimization is published on our site.
Assembled prototype after topological optimization. Video of manipulator operation
The final design after topological optimization — characteristic "organic" shapes of the stiffening ribs.
Final design and concept
The final design combines all the developments: semi-SCARA kinematics, paired servos for backlash compensation, an optimized BOM, and topologically optimized elements.
Tabletop installation of 4 manipulators
Manipulator without the cover
Use on a mobile platform
In summary: what we got
Key decisions that made it possible to lower the cost:
- Budget servo drives (Feetech STS3215 or Feetech STS3250, $15–40 per unit) instead of expensive Dynamixel.
- Semi-SCARA kinematics — reduction of the load on the drives at the expense of the horizontal operation of part of the links.
- Paired servos with preload — backlash compensation without expensive zero-backlash gearboxes.
- Radical simplification of the BOM — from 46 to 6 custom parts.
- Topological optimization — reduction of stresses by 3.7 times with a mass increase of only 22%.
Cost of components for one manipulator: ~$900.
- GitHub: github.com/roboninecom
- Telegram: t.me/robo_9
Top comments (0)