Design Book › Worked example: the SO-101 arm

Worked example: the SO-101 arm

Result. At full reach (0.411 m), arm level, no payload, the shoulder lift must hold 0.910 N·m (9.28 kg·cm).

The robot is SO-ARM101. The arithmetic is in tools/book_calcs/so101_torque.py; its output is tools/book_calcs/out/so101_torque.md.

Parts list and cost

The bill of materials has 28 rows: 14 priced, 14 unpriced. All priced rows are in USD and cite the project README.

Parts list and cost
ItemFollower qtyPair qtyUnit USDFollower USDPair USD
STS3215-C001, 7.4 V, 1/3456713.8983.3497.23
STS3215-C044, 7.4 V, 1/1910213.89027.78
STS3215-C046, 7.4 V, 1/1470313.89041.67
Motor control board1210.6010.6021.20
Power supply1210.0010.0020.00
USB-C cable, 2 pcs117.007.007.00
Table clamp (2 pcs for follower, 4 pcs for pair)115.00, 9.005.009.00
Screwdriver set116.006.006.00
Total121.94229.88

The totals are sums of the priced BOM rows above.

Parts list and cost
GroupFollower USDSharePair USDShare
Servos83.3468.3%166.6872.5%
Tools and clamps11.009.0%15.006.5%
Motor control boards10.608.7%21.209.2%
Power supplies10.008.2%20.008.7%
Cables7.005.7%7.003.0%

Priced and not priced:

The $13.89 unit price is one Alibaba listing. Other rows for the same family:

Parts list and cost
PartUSDVendorWhere
12 V STS321513.89AlibabaLeKiwi parts list
12 V STS321514.00, 16.00Wowrobo, AlibabaXLeRobot parts list
7.4 V STS321515.00AlibabaLeKiwi parts list
STS3215-C001, 7.4 V14.65 (97.72 CNY)TaoBaoSO-ARM100 README
STS3215-C001, 7.4 V18.82 (2,980 JPY)Akizuki DenshiSO-ARM100 README

At the Akizuki price of $18.82, the six follower servos add 6 x (18.82 - 13.89) = $29.58, for $151.52. This site shows no store price for the 12 V part beyond the parts lists above. Price the servo you will buy, not the BOM row. See Servos.

Servo variants

One family name covers six catalog rows. Their values differ [2]. The table shows the four rows with values this site can show. All of them come from the SO-ARM100 README.

Servo variants
PartRatioVoltageStallWhere in the README
C001345:17.4 V16.5 kg·cm at 6 Vparts list; text on the 7.4 V version
C044191:17.4 Vnot shownparts list
C046147:17.4 Vnot shownparts list
C018not shown12 V30 kg·cmtext on the 12 V version

The other two rows are C047 and a 7.4 V part with no suffix. Their values come from a datasheet copy and a retailer, and this site shows neither part. Rated torque, no-load speed, supply range and current are not in the table. For these servos the database holds them only from the maker wiki and seller tables, which this site cannot show. Take them from the sheet of the servo you buy.

Three cautions from the source records:

Which voltage the sources give

Which voltage the sources give
SourceLeaderFollowerRow
SO-ARM100 README (so-arm100)7.4 V: 1 C001, 2 C044, 3 C0467.4 V: 6 C001bom rows; leader count is pair minus follower
NVIDIA SO-101 course (nvidia-so101-course)5 V supply12 V supply[3]
LeRobot docs (lerobot-docs)different gear ratios; voltage not in the quote6 STS3215 at 1/345; voltage not in the quote[4]
LeKiwi BOM, 12 V set (lekiwi)12 V; 12 servos in the pair, one row, part id 12 Vsame rowLeKiwi
LeKiwi BOM, 5 V set (lekiwi)5 V build; 12 servos in the pair, one row, part id 7.4 Vsame rowLeKiwi
XLeRobot BOM (xlerobot)no leader row12 V; 12 servos under armsXLeRobot

A 5 V leader runs 7.4 V servos well below their nominal voltage. The database holds a supply range for C001 and C044 only from the maker wiki and a seller, and none for C046. Check the minimum voltage on each servo's sheet before you build a 5 V leader. Each arm has its own control board and supply in the pair rows, so the two voltages do not share a bus. Label them [3].

Static torque check

Method

  1. Read link masses and frames from the upstream simulation model robotstudio_so101/so101.xml. Every mass and length here comes from that simulation model, not from a weighed arm. The database holds only the base link mass for this robot.
  2. Walk the MJCF tree with forward kinematics. Positions are local to the parent [5].
  3. Find full reach: the pose with the largest horizontal distance from the shoulder lift axis to the gripperframe site. The result is 0.4110 m, with the gripper frame level with the axis (0.1167 m against 0.1166 m above the base). Joint angles in rad: shoulder_lift 1.3264, elbow_flex -1.2878, wrist_flex 0.0365, others 0.
  4. At each joint, sum m x g x arm over all bodies beyond it, where arm is the horizontal distance from the joint axis to the body's centre of mass [6]. g is 9.80665 m/s², the same constant the catalog uses to convert kg·cm to N·m.
  5. Repeat over the joint ranges: a 25 x 25 x 25 x 5 x 2 grid on lift, elbow, wrist flex, wrist roll and gripper, then a local search from the best point [7].

Cross-check. MuJoCo 4.0.0 on the same file and pose gives gravity forces of 0.9097, 0.4834, 0.1298, 0.0068 and 0.0035 N·m at the five loaded joints, with its default g of 9.81. The script gives 0.9096, 0.4834, 0.1299, 0.0071 and 0.0035 N·m. See tools/book_calcs/mujoco_crosscheck.py.

The base is taken as level, so shoulder_pan has a vertical axis and carries no gravity load. The camera_mount body has no inertial element. Its mass is 0.012 kg from the mesh geom, plus 0.00144 kg for a box with no mass attribute, at the water density MuJoCo assumes [8]. The parts list has no camera; the model does.

Arithmetic at full reach

Shoulder lift:

Arithmetic at full reach
BodyMass (kg)Arm (m)m x g x arm (N·m)
upper_arm0.10300.08920.0901
lower_arm0.10400.21410.2184
wrist0.07900.28950.2243
gripper0.08700.33710.2876
camera_mount0.01340.34950.0461
moving_jaw0.01200.36660.0431
Sum0.39840.9096

Elbow flex:

Arithmetic at full reach
BodyMass (kg)Arm (m)m x g x arm (N·m)
lower_arm0.10400.09810.1001
wrist0.07900.17350.1344
gripper0.08700.22110.1886
camera_mount0.01340.23350.0308
moving_jaw0.01200.25060.0295
Sum0.29540.4834

Wrist flex:

Arithmetic at full reach
BodyMass (kg)Arm (m)m x g x arm (N·m)
wrist0.07900.03850.0298
gripper0.08700.08610.0734
camera_mount0.01340.09850.0130
moving_jaw0.01200.11560.0136
Sum0.19140.1299

Load at each joint

Load at each joint
JointFull reach (N·m)Full reach (kg·cm)Same, no camera mount (N·m)Worst over all poses (N·m)
shoulder_pan0.0000.000.0000.000
shoulder_lift0.9109.280.8630.910
elbow_flex0.4834.930.4530.484
wrist_flex0.1301.320.1170.130
wrist_roll0.0070.070.0000.009
gripper0.0040.040.0040.004

Full reach is the worst pose found for the lift and wrist flex, and within 0.001 N·m of the worst for the elbow. Wrist roll peaks in another pose, at 0.009 N·m. The gripper line is the jaw's weight only. It says nothing about grip force.

Compared with the catalog

Each cell is limit divided by load. Under 1.00 fails that limit. Both limits are the README stall figures [2].

Compared with the catalog
LimitN·mshoulder_liftelbow_flexwrist_flex
C001 stall, 6 V (README)1.6181.783.3512.46
C018 stall, 12 V (README)2.9423.236.0922.66

The script also prints rows for rated torque, for the maker wiki's stall values and for a bench hold test. This site cannot show those inputs, so those rows are left out here. The rated comparison needs the maker's rated torque: take it from the sheet of the servo you buy and divide it by the loads in the table above. A result under 1.00 fails.

Wrist roll and gripper load is under 0.01 N·m. They have over 50 times margin on both limits above.

Payload at the gripper frame, full reach. Payload = (limit - load) / (g x tip arm). The tip arm is 0.4107 m at the shoulder lift, 0.2947 m at the elbow, 0.1597 m at the wrist. The shoulder lift governs in every row.

Compared with the catalog
LimitPayload (kg)
C001 stall, 6 V (README)0.176
C018 stall, 12 V (README)0.505

Both rows use stall, a momentary limit, so they are upper bounds. For the payload you can hold, put your servo's rated torque in the formula: (rated - 0.910) / (9.80665 x 0.4107) kg at the shoulder lift. A negative result means no payload at full reach.

The database holds one payload claim for this arm: 150 g rated, stated by Robonine, not by The Robot Studio. The record gives no pose or voltage, so it does not test these numbers.

Torque margin at the shoulder lift. TM = T_avail / (FS x T) - 1 [9]. This torque comes from analysis, so the analysis-factor guideline gives 1.50 on fixed and 3.00 on variable torque [10]. The guideline does not say which class gravity falls in, so both are shown. The factors come from a NASA mechanism standard (nasa-std-5017). Read them here as a method, not a requirement for a bench arm.

Compared with the catalog
T_availFS = 1.00FS = 1.50FS = 3.00
C001 stall, 6 V (README)+0.78+0.19-0.41
C018 stall, 12 V (README)+2.23+1.16+0.08

Stall is a momentary limit, so these rows are the best case. The margin on rated torque uses the same formula with your servo's rated torque as T_avail. With FS = 1.50 it is zero or more only if the rating is at least 1.50 x 0.910 = 1.365 N·m (13.9 kg·cm).

Duty. The guideline compares rated torque with torque over the duty cycle [11], [12], and full reach is rarely held. If the arm spends a fraction d of the cycle at 0.910 N·m and the rest at zero load, then M_RMS = sqrt(d) x 0.910, so d <= (rated / 0.910)^2. For example, a servo rated at half the load can hold full reach for at most a quarter of the cycle. A rating of 0.910 N·m or more passes at any duty.

This is an upper bound, because no real cycle rests at zero load. The database has no task duty cycle for this arm. Load also falls with the cosine of the tilt from level. With a servo rated under 0.910 N·m, the load stays over the rating until the arm tilts past arccos(rated / 0.910) from level.

Measured backlash and heat

The database holds bench results for the 12 V C018 from two records: Robonine (robonine, a bench write-up) and SO-ARM100 issue 134 (so-arm100-issues, a user bench table). The catalog's source note says the issue links to the Robonine write-up. Read them as one test series, not two independent tests. The database gives no sample size. This site cannot show the values of either record, so the hold torque, no-load speed, repeatability and temperature rise are not shown. Two guidelines state results from these tests. This section uses only those.

How to read the results:

Measured backlash and heat
Angleshoulder_lift (0.4107 m)elbow_flex (0.2947 m)wrist_flex (0.1597 m)Sum of the three
0.5° (datasheet limit)3.58 mm2.57 mm1.39 mm7.55 mm
1° (measured)7.17 mm5.14 mm2.79 mm15.10 mm

The sum is the worst case in the arm plane, with all three joints at one end of their play. Compare it with the encoder step: 360° divided by the steps per turn on your servo's sheet. The database holds the STS3215 encoder resolution only in a datasheet row that this site cannot show. The database does not say where the encoder sits relative to the gears, so it does not say whether the encoder sees this play. In simulation, model the play on purpose [16], [17].

Upstream simulation against the catalog

Left column: robotstudio_so101/so101.xml and its README, from the MuJoCo Menagerie. These are simulation values, not maker data.

Upstream simulation against the catalog
ItemUpstream simulationCatalogReading
Actuator force limitforcerange ±2.94 N·m, one class for all six jointsC018 stall 2.942 N·m (README); C001 stall 1.618 N·m at 6 V (README)Equals the 12 V stall to three figures. It is 1.82 times the C001 stall at 6 V. Rated torque is below stall, so the limit is above any rated torque too. See the effort-limit guideline [18].
Earlier force limitREADME: "Use default forcerange for actuators rather than 3.35 N/m"3.35 N·m = 34.16 kg·cmNo catalog row equals this. The basis is not stated.
Speed limitnone on the actuatorsnot shownThe database holds STS3215 speeds only from sources this site cannot show. The model has no torque-speed curve [19].
Backlashclass backlash spans ±0.0087267 rad (±0.5°); file comment says "not added to any joints for now"at most 0.5° (datasheet); about 1° (bench) [13]Running model has zero backlash. The sources do not say whether "0.5°" is end to end or per side.
Armature0.028 kg·m²no rotor inertia for any STS3215 row; published fitted armature exists [20], [21]The armature formula cannot be checked [22]. Implied rotor inertia at 345:1 is 0.028 / 345² = 2.35e-7 kg·m², derived, not sourced.
Damping, frictionloss0.60, 0.052no part fact; published fits exist [23], [24], [21]No part value to compare. The fits belong to other models.
Position gainskp 998.22, kv 2.731; file comment: not a 1-to-1 mapping, assumes servo gain 16no part fact; published fitted stiffness exists [25], [21]Treat as tuned, not measured.
Total mass0.6454 kg, including the camera mountnot shown: the database holds STS3215 mass only from the maker wiki and sellersNo weighed arm in the database, so the total-mass check stays open [26].
Leader armno modelleader has mixed ratiosThe leader armature check stays open [4]. No leader torque check is possible.
SO-ARM100 sibling modelsMenagerie trs_so_arm100: ±3.5 N·m. Repo so100.urdf: effort="35", velocity="1"C018 stall 2.942 N·m (README)The catalog reads the URDF effort as N·m, which is 11.9 times the 12 V stall. The URDF speed limit has no public no-load speed to compare with. Three related models use three force limits [18].

Checks the upstream file passes:

The practical problem is the voltage. A policy trained in this model can use 2.94 N·m at any joint. The README stall of the 7.4 V servo is 1.618 N·m at 6 V, well under that. A 12 V servo reaches it only for a moment, at stall. Lower forcerange to match your build before you trust a transfer.

What is settled and what is not

Settled, on the records in this database:

Not settled:

Guidelines this example exercises

Guidelines this example exercises
GuidelineResult here
A servo's stall torque is its maximum momentary output and differs from its continuous rating. Real-world output is generally closer to the performance-graph measurements. [1]Open: the check needs the maker's rated torque, which this site cannot show. The shoulder lift needs 0.910 N·m at full reach.
The STS3215 7.4 V version is rated at 16.5 kg·cm stall torque at 6 V; the 12 V version is rated at 30 kg·cm. [2]Applied.
The selected motor's stall torque should usually exceed the load peak torque. [36]Stall exceeds the static load, at the output. Motor shaft values are not in the database.
For NASA spaceflight mechanisms, calculate torque margin as available torque over factored resisting torques minus one; zero or more meets the requirement [9]; For NASA-STD-5017B theory-or-analysis torque data, use minimum factors of 3.00 on variable, 1.50 on fixed, and 1.25 on acceleration torque [10]; Keep static torque margin above zero over the full range of motion. [7]Margin computed on the two README stall figures. The margin on rated torque needs the servo's sheet.
For start-stop or varying cycles, keep the motor nominal torque above the RMS load torque [11]; Compute cycle RMS torque as the square root of the sum of time times torque squared, divided by total cycle time including rest. [12]Gives the duty bound as a formula. The rated torque comes from the servo's sheet.
In a Robonine STS3215 test, issues appeared at 71°C and no built-in thermal protection was apparent, so manual shutdown is needed at that temperature. A later page comment says firmware protection cuts torque near 70°C [14]; Urs et al. (3D-printed actuators) list 20 s peak, 2 s peak and continuous torque separately and note the Mini Cheetah peak duration was not provided. [37]The 71 °C result, and stall figures with no duration.
A bench test of the STS3215 measured about 1° of encoder backlash against a ≤0.5° datasheet value. [13]Measured value used for tip error.
Do not assume current maps to torque at gear ratios well above 10; backlash, vibration and friction can no longer be ignored. [15]Applies at 345:1.
Include gravity in the joint torque budget from the equations of motion, tau = M ddq + c + g. [6]Gravity torque over the workspace, by script.
Set armature for a geared actuator from rotor inertia and ratio, as J times G squared, where G is motor speed over output speed [22]; The SO-101 leader uses several gear ratios, while the follower uses 1/345 gearing [4]; Model MuJoCo backlash with two coaxial hinges, and put armature on the driven one, or the joint-space inertia matrix is singular [16]; Set each simulated joint's maximum force from the actuator specification when one exists [18]; Derive each joint's effort limit from its own motor and gearbox; a single large value across the robot is not acceptable [33]; Use DCMotor when the actuator model needs a four-quadrant torque-speed envelope rather than a fixed effort clip [19]; Keep position command ranges inside joint ranges, except for a gripper that must close past pad contact [27]; Set damping and a force bound on every MuJoCo position actuator, because kv defaults to zero and force is unlimited. [28]Lint of the upstream model.
Treat MJCF positions and orientations as local to the parent, not as world values [5]; Native MJCF sensors do not require massless dummy mounting bodies [32]; Before any tuning, check mesh scale and total mass to catch a millimetre-for-metre mistake [26]; MJCF angles default to degrees, while URDF input is always parsed as radians [29]; Do not rely on inferred mass for hollow printed shells or dense metal links, because MuJoCo assumes water density. [8]Model reading and units.
Model friction, backlash and thermal effects where they matter, because typical motor models lack them [17]; Use elliptic cones with impratio about 10 for MuJoCo manipulation and 10-100 for legged robots when slip matters [30]; Set MuJoCo integrator, cone and solref explicitly in each new model, because their internal defaults are not the recommended values. [31]Simulation setup.
Label the SO-101 leader and follower supplies, because they differ at 5 V and 12 V and must not be swapped. [3]Label the two supplies.
Check supply limits before motor selection, since battery and amplifier limits can set the available peak current. [34]Open: no supply rating in the BOM.

For the general method, see chapter 4.

Sources

Each guideline in this chapter with its source, the date it was read and the exact words of the source. Level: must means breaking it causes failure or harm, should is good practice, guide is a rule of thumb.

  1. A servo's stall torque is its maximum momentary output and differs from its continuous rating. Real-world output is generally closer to the performance-graph measurements. Level: guide. emanual.robotis.com 2026-11-08
  2. The STS3215 7.4 V version is rated at 16.5 kg·cm stall torque at 6 V; the 12 V version is rated at 30 kg·cm. Level: guide. github.com 2026-11-08“The 12V version has a stall torque of 30kg.cm.”
  3. Label the SO-101 leader and follower supplies, because they differ at 5 V and 12 V and must not be swapped. 5 and 12 V Level: must. docs.nvidia.com 2026-11-08
  4. The SO-101 leader uses several gear ratios, while the follower uses 1/345 gearing. Level: guide. huggingface.co 2026-11-08
  5. Treat MJCF positions and orientations as local to the parent, not as world values. Level: must. mujoco.readthedocs.io 2026-11-08
  6. Include gravity in the joint torque budget from the equations of motion, tau = M ddq + c + g. Level: should. hades.mech.northwestern.edu 2026-11-08
  7. Keep static torque margin above zero over the full range of motion. Level: must. standards.nasa.gov 2026-11-08“[DDMR 12] Static torque margin per Equation 4-1 and Table 1 shall be greater than zero within”
  8. Do not rely on inferred mass for hollow printed shells or dense metal links, because MuJoCo assumes water density. Level: must. mujoco.readthedocs.io 2026-11-08
  9. For NASA spaceflight mechanisms, calculate torque margin as available torque over factored resisting torques minus one; zero or more meets the requirement. Level: must. standards.nasa.gov 2026-11-08“The required reserve torque is included in the equation in the form of the safety factors, so a”
  10. For NASA-STD-5017B theory-or-analysis torque data, use minimum factors of 3.00 on variable, 1.50 on fixed, and 1.25 on acceleration torque. 3.0 Level: must. standards.nasa.gov 2026-11-08“Theory or analysis 3.00 1.50 1.25”
  11. For start-stop or varying cycles, keep the motor nominal torque above the RMS load torque. Level: must. maxonmotor.com 2026-11-08
  12. Compute cycle RMS torque as the square root of the sum of time times torque squared, divided by total cycle time including rest. Level: guide. maxonmotor.com 2026-11-08
  13. A bench test of the STS3215 measured about 1° of encoder backlash against a ≤0.5° datasheet value. Level: guide. github.com 2026-11-08
  14. In a Robonine STS3215 test, issues appeared at 71°C and no built-in thermal protection was apparent, so manual shutdown is needed at that temperature. A later page comment says firmware protection cuts torque near 70°C. Level: should. robonine.com 2026-11-08
  15. Do not assume current maps to torque at gear ratios well above 10; backlash, vibration and friction can no longer be ignored. Level: guide. manipulation.csail.mit.edu 2026-11-08
  16. Model MuJoCo backlash with two coaxial hinges, and put armature on the driven one, or the joint-space inertia matrix is singular. Level: must. mujoco.readthedocs.io 2026-11-08
  17. Model friction, backlash and thermal effects where they matter, because typical motor models lack them. Level: should. docs.nvidia.com 2026-11-08
  18. Set each simulated joint's maximum force from the actuator specification when one exists. Level: should. docs.omniverse.nvidia.com 2026-11-08
  19. Use DCMotor when the actuator model needs a four-quadrant torque-speed envelope rather than a fixed effort clip. Level: guide. isaac-sim.github.io 2026-11-08“Same PD torque, clipped to a four-quadrant torque-speed envelope.”
  20. The Playground v1.0.0 SO-101 armature values identified from real teleop data are 0.1088 shoulder pan, 0.1472 shoulder lift, 0.1224 elbow, 0.1133 wrist flex, 0.1146 wrist roll and 0.1985 gripper. Level: guide. raw.githubusercontent.com 2026-11-09“# real teleop data. ARMATURE = { "shoulder_pan": 0.1088, "shoulder_lift": 0.1472, "elbow_flex": 0.1224, "wrist_flex": 0.1133, "wrist_roll": 0.1146, "gripper": 0.1985,”
  21. The gtc_berlin fit gives the SO-101 shoulder pan stiffness 48.93682087435154, damping 2.309705389862818 and armature 0.058228342618297214. Level: guide. raw.githubusercontent.com 2026-11-09“stiffness: 48.93682087435154 damping: 2.309705389862818 armature: 0.058228342618297214”
  22. Set armature for a geared actuator from rotor inertia and ratio, as J times G squared, where G is motor speed over output speed. Level: should. docs.omniverse.nvidia.com 2026-11-08
  23. The Playground v1.0.0 SO-101 identified damping runs from 17.5096 at shoulder pan to 34.1633 at the gripper. Level: guide. raw.githubusercontent.com 2026-11-09“"shoulder_pan": 17.5096, "shoulder_lift": 28.0620, "elbow_flex": 25.6669, "wrist_flex": 25.4543, "wrist_roll": 23.1058, "gripper": 34.1633,”
  24. The Playground v1.0.0 SO-101 friction loss, identified from real teleop data, is 0.2517 at the gripper and near zero on arm joints. 0.2517 Level: guide. raw.githubusercontent.com 2026-11-09“# Per-joint frictionloss identified from real teleop data. FRICTIONLOSS = { "shoulder_pan": 0.0001, "shoulder_lift": 0.0, "elbow_flex": 0.0, "wrist_flex": 0.0, "wrist_roll": 0.0, "gripper": 0.2517,”
  25. The Playground v1.0.0 SO-101 position gains, identified from real teleop data, set arm stiffness from 218.4066 at wrist roll to 265.0958 at shoulder lift. Level: guide. raw.githubusercontent.com 2026-11-09“# Position servo PD gains identified from real teleop data. STIFFNESS = { "shoulder_pan": 251.5896, "shoulder_lift": 265.0958, "elbow_flex": 243.9844, "wrist_flex": 223.2804, "wrist_roll": 218.4066,”
  26. Before any tuning, check mesh scale and total mass to catch a millimetre-for-metre mistake. Level: must. mujoco.readthedocs.io 2026-11-08
  27. Keep position command ranges inside joint ranges, except for a gripper that must close past pad contact. Level: must. mujoco.readthedocs.io 2026-11-08
  28. Set damping and a force bound on every MuJoCo position actuator, because kv defaults to zero and force is unlimited. Level: should. mujoco.readthedocs.io 2026-11-08
  29. MJCF angles default to degrees, while URDF input is always parsed as radians. Level: guide. mujoco.readthedocs.io 2026-11-08
  30. Use elliptic cones with impratio about 10 for MuJoCo manipulation and 10-100 for legged robots when slip matters. 10 impratio Level: should. mujoco.readthedocs.io 2026-11-08
  31. Set MuJoCo integrator, cone and solref explicitly in each new model, because their internal defaults are not the recommended values. Level: should. mujoco.readthedocs.io 2026-11-08
  32. Native MJCF sensors do not require massless dummy mounting bodies. Level: guide. mujoco.readthedocs.io 2026-11-08
  33. Derive each joint's effort limit from its own motor and gearbox; a single large value across the robot is not acceptable. Level: should. isaac-sim.github.io 2026-11-08“choosing both from the motor and gearbox rather than using one oversized value for the whole robot.”
  34. Check supply limits before motor selection, since battery and amplifier limits can set the available peak current. Level: must. maxonmotor.com 2026-11-08
  35. The SO-101 leader uses mixed gearing; its shoulder pan motor, for one, is geared 1/191, so check the table for each joint. Level: guide. raw.githubusercontent.com 2026-11-09“| Base / Shoulder Pan | 1 | 1 / 191 |”
  36. The selected motor's stall torque should usually exceed the load peak torque. Level: should. maxonmotor.com 2026-11-08
  37. Urs et al. (3D-printed actuators) list 20 s peak, 2 s peak and continuous torque separately and note the Mini Cheetah peak duration was not provided. Level: guide. arxiv.org 2026-11-08