Actuators: quasi-direct drive, torque, speed and heat

Stage 2 · Mechanics · ~40 min · Before this: 01 Degrees of freedom, 03 System architecture · numbers from the repository as of 2026-11-02

Every joint of a humanoid is only as strong, fast and gentle as the actuator behind it. This page explains what an electric actuator can and cannot do (how much torque, at what speed, for how long before it overheats), and why Marvin Bot uses four sizes of low-ratio RobStride actuators.

You will learn

• How current becomes torque and speed becomes voltage in a brushless motor (torque constant and back-EMF), and why that gives a torque-speed envelope.
• What a gearbox trades: torque up by N, speed down by N, reflected inertia up by N², and what “quasi-direct drive” means.
• Why rated (continuous) torque is a heat limit, not a strength limit, and how the heat sink and the duty cycle change it.
• Why each RobStride model sits at each of Marvin Bot’s joints, and where the knee runs out of margin on stairs.

The idea in one picture

Actuators are the robot’s muscles: every joint from page 01 has one, and every balancing move from page 02 has to come out of them. You already know their three limits from your own arm. You can lift a heavy shopping bag for a moment, but not for an hour; you can swing an empty arm fast, but not a loaded one; and if you hold the bag out for long, the muscle gets tired and hot. A motor is the same, except that “tired” means “the copper inside is getting too hot”.

Two units first. Torque is turning force, measured in newton-metres (N·m): holding a 1 kg bag 10 cm from a joint takes about 1 N·m, and Marvin Bot’s knee holds about 47 N·m on a high stair step. Joint speed is in radians per second (rad/s; 1 rad/s is about 9.5 revolutions per minute, rpm).

An actuator here means a brushless electric motor (one with no rubbing contacts inside, so it lasts and runs cool), a small gearbox and a motor driver (the electronics that feed it) in one housing. The motor turns electric current into torque; the gearbox multiplies that torque and divides the speed; the driver decides, a few thousand times a second, how much current to push. Three limits shape everything it can do:

• Current limit (the flat top of the envelope): the driver and the magnets allow only so much current, so only so much peak torque.
• Voltage limit (the falling edge): a spinning motor generates a voltage against its supply, the back-EMF. Near the top speed there is no voltage left to push current, so torque falls to zero at the no-load speed.
• Heat limit (the dashed line): current heats the copper windings. The torque the motor can hold forever without overheating is the continuous (rated) torque, typically a third of the peak.

So there are two different “how strong?” numbers: the peak torque (for a moment) and the rated torque (for ever). Most design questions on this page come down to which of the two a task needs. The first playground draws these limits for Marvin Bot’s actuators and lets you ask “can this joint do that?”. The second follows the winding temperature through a stair climb. The third shows why each model sits where it does.

Playground 1: torque-speed explorer

Start here: press Marvin Bot today, then try the other presets. The two “What the joint needs” sliders place a dot: the torque and speed a task asks of the joint.

How to read this picture: speed runs left to right, torque bottom to top. The outer (blue) line is the most the actuator can give for a moment; the green area is what it can give for ever without overheating. A dot inside green is fine; inside the outer line but above green is fine only briefly (a heat problem); outside the outer line is impossible. The orange band shows the speed range the motor’s command format allows (explained under “Driver modes” below).

Presets

Marvin Bot today: knee on an 18 cm step Knee swing while walking Parked 1.5:1 knee stage 50:1 high-ratio knee Nearly empty battery

Actuator

RobStride model RS04 (120 N·m peak, hips and knee)RS03 (60 N·m, hip yaw, ankles, waist, shoulder pitch)RS10P (42 N·m, shoulder roll and yaw, elbow)RS00 (14 N·m, wrist yaw)

Total gear ratio

Rescales the chosen motor as if it had a different gearbox (plus an extra stage). Native: 9:1 (RS00 10:1).

Bus voltage

The 48 V curve is always drawn for comparison. The battery runs from about 42 V (nearly empty) to 54.6 V (full).

What the joint needs

Joint torque

How hard the joint must turn (N·m). 47 N·m = the knee holding the robot on an 18 cm step.

Joint speed

How fast the joint must turn (rad/s; 1 rad/s ≈ 57° per second). A knee swinging in a walk reaches about 8 rad/s.

Show the simulator's motor model

Isaac Lab's DC-motor model as Marvin Bot configures it: peak torque falling in a straight line to zero at the no-load speed.

Torque at the joint versus joint speed

–

Peak / rated torque

–

No-load speed at this voltage

–

Torque available at this speed

–

Motor current (rms)

–

Heat vs rated

–

Mechanical power

–

Reflected inertia

–

Felt at the sole (knee, 0.375 m)

–

peak envelope, this voltagepeak envelope, 48 Vcontinuous (rated) regionsimulator's straight-line modelMIT-mode speed range

Try this

1. Press Marvin Bot today. The point (47 N·m, nearly still) is far inside the peak envelope but above the green continuous region: a heat problem, not a strength problem. Now press Parked 1.5:1 knee stage: the same point drops inside the continuous region, but the top speed falls from about 21 to 14 rad/s.
2. Drag the gear ratio to 50. Torque looks wonderful, but watch Felt at the sole: the motor’s spinning rotor now feels like about 7 kg bolted to the foot on impact, against about 0.2 kg at 9:1.
3. Press Knee swing while walking and drag the voltage from 54 down to 36 V. The top of the envelope does not move; only the speed edge does.
4. Switch on the simulator’s motor model (violet). Compare it with the manual’s curve: the straight line is much more pessimistic at mid speeds. That is a deliberate safety margin in Marvin Bot’s training, and one of the things the bench test should calibrate.

From first principles

Current makes torque: the torque constant

Inside a brushless motor, permanent magnets on the rotor (the part that spins) sit next to copper coils on the stator (the fixed part). Current in a coil in a magnetic field feels a force; arranged around a circle, that force is a torque. The driver keeps the current always pointing the most useful way (this is called field-oriented control, FOC), so torque is simply proportional to current:

τ = K_t · I

τ: torque (N·m). I: phase current (A rms). K_t: torque constant (N·m/A). RobStride quotes K_t at the output of the gearbox: 2.1 N·m/A for the RS04, 2.36 for the RS03, 1.09 for the RS06 and 1.48 for the RS00.^[2][3][4][5]

Check it against the RS04 manual: rated current 27 A peak = 27/√2 = 19.1 A rms, times 2.1 N·m/A = 40 N·m, exactly the rated torque. At the peak current (90 A peak = 63.6 A rms) the same sum gives 134 N·m, but the manual’s peak is 120 N·m: at high current the iron saturates and K_t drops by about 10%.^[2] The playground’s current readout uses this measured curve.

Speed makes voltage: back-EMF

The same magnets passing the same coils generate a voltage proportional to speed, the back-EMF. The supply must overcome it to push current in, so a motor on a fixed voltage has a top speed where the back-EMF eats the whole supply: the no-load speed.

ω₀ ≈ V_bus / (√2 · k_e) (motor speed, then ÷ N at the output)

ω₀: no-load speed. V_bus: battery voltage (V). k_e: back-EMF constant, line-to-line, in V rms per 1000 rpm of the rotor. √2 turns rms into the peak a 48 V bus can match.

The RS03 manual gives 17 V rms per 1000 rpm.^[3] At 48 V: 48/(1.414 × 17) = 2.0 thousand rpm at the rotor, ÷ 9 = 222 rpm at the output. The manual measures 195-200 rpm; the gap is resistance and losses. Two consequences matter for a robot: speed scales with battery voltage (a full 54.6 V pack is about 14% faster than 48 V, a nearly empty 42 V pack about 12% slower), and torque falls as speed approaches the top.

Go deeper: how the two constants are linked, and field weakening

The two constants are the same physics seen from two sides: for an ideal three-phase motor K_t = √3 · k_e (in SI units), which for the RS03 predicts 2.53 N·m/A against the quoted 2.36, consistent with a gearbox around 93% efficient.

Some drivers can push past the top speed with field weakening: an extra current component that partly cancels the magnets’ field, lowering the back-EMF at the cost of more heat and less torque.^[13] RobStride drivers have a field-weakening current parameter (I_FW_MAX), set to 0 (off) by default.^[2] Marvin Bot does not use it, and the envelope here assumes it is off.

The torque-speed envelope and power

Put the two limits together and you get the envelope in the playground: a flat top at the peak torque, then a falling edge down to the no-load speed. The manuals publish this “T-N curve” at 48 V; the playground uses those curves, digitised by hand from the manuals’ plots.^[2][3] Mechanical power is torque times speed, P = τ · ω (watts, with ω in rad/s). The RS04’s curve passes about 78 N·m at 150 rpm (15.7 rad/s): 1.2 kW of short-burst output from a 1.42 kg actuator.

A physics simulator needs a simpler rule. Isaac Lab’s DCMotor model clips torque to a straight line from the “saturation” torque at standstill to zero at the maximum speed.^[11] Marvin Bot sets the saturation torque equal to the peak torque (sim/isaaclab/marvin_bot/robot_cfg.py), which is the violet triangle in the playground: at 95 rpm the RS04 manual shows the full 120 N·m, while the straight line allows 63 N·m. Training against a weaker motor is safe in one direction (the policy never asks for torque that does not exist) but can make it slower and more cautious than the hardware needs to be.

The gearbox: torque × N, speed ÷ N, inertia × N²

A gearbox with ratio N (for example 9:1: the motor turns 9 times for each turn of the joint) multiplies torque by N and divides speed by N, like the low gears on a bicycle. Power stays the same, minus friction. So why not use a large ratio and a tiny motor? Because of three side effects:

• Reflected inertia grows with N². The rotor spins N times faster than the joint, so to accelerate the joint you must accelerate the rotor N times as much, and the rotor’s kinetic energy is N² times larger. Seen from the joint, the rotor’s inertia J_rotor looks like J_refl = N² · J_rotor.^[10] In a foot strike, this inertia must be stopped instantly, so it acts like extra mass at the foot, and the impact force grows with it.^[6]
• Backdrivability falls. Gear friction scales roughly with N, and at high ratios you cannot turn the output by hand. A backdrivable actuator lets the leg “give” on impact, and lets the driver estimate the joint torque from the motor current alone, without a torque sensor.^[6]
• Impact tolerance falls. A gear tooth must stop the rotor’s reflected inertia in an impact; high-ratio gears are where humanoid legs break.

The playground reports reflected inertia from the armature values Marvin Bot’s simulation uses (0.03 kg·m² for an RS04 joint, 0.02 for RS03, 0.00625 for RS10P (RobStride’s figure at the output^[15]), 0.005 for RS00; the others placeholders until the bench test) and divides it by the square of the knee-to-sole distance (0.315 m shin + 0.06 m ankle) to show the “mass” a foot strike feels.

Quasi-direct drive

“Quasi” means “almost”: almost no gearbox. The MIT Cheetah group showed that a large-diameter motor (lots of torque per amp, because torque grows with the rotor’s radius) with a single low-ratio planetary stage (a compact gearbox with small “planet” gears running round a central “sun” gear), about 6:1 to 10:1, keeps the benefits of direct drive (low reflected inertia, backdrivability, current-based force control) while getting enough torque. They called it proprioceptive actuation: the actuator senses force from its own current.^[6][7] The Mini Cheetah made it cheap and modular.^[8] This is now called quasi-direct drive (QDD), and RobStride’s actuators are of this kind: 9:1 or 10:1 planetary gearboxes with a FOC driver on the back.

Other gearboxes

Type	Typical ratio	Good at	Costs
Planetary (QDD)	4:1-10:1 per stage	backdrivable, efficient, low reflected inertia, cheap	some backlash; one stage limits the ratio
Harmonic (strain-wave)	30:1-160:1	no backlash, compact, precise	hard to backdrive, high reflected inertia, flexspline sensitive to shock loads[10]
Cycloidal	10:1-100:1	load shared over many lobes, shock tolerant, compact	more parts; friction rises with ratio. Berkeley Humanoid Lite uses 3D-printed cycloidal gearboxes because they tolerate printed-plastic tolerances better than planetary gears[9]
Linkage or belt stage	1.2:1-3:1	moves the motor and adds a little ratio	parts, some compliance; the parked 1.5:1 knee option is this kind

Heat: copper loss, thermal resistance and time constants

Current through the windings’ resistance R makes heat, the copper loss (Joule heating):

P_cu = 3 · I² · R_phase ∝ τ²

R is the winding’s electrical resistance per phase (ohms) and the 3 counts the motor’s three phases (its three sets of coils); ∝ means “is proportional to”. Torque is proportional to current, so heat goes with torque squared: 118% of rated torque makes 1.39 times the rated heat; 3 times rated (the peak) makes 9 times. Copper’s resistance also rises 0.39% per kelvin (a temperature step of 1 K is the same as 1 °C)^[14], so a hot winding (145 °C) makes about 47% more heat for the same current than a cold one.

Think of a kettle on a low flame: it warms quickly at first, then levels off where the heat put in equals the heat lost to the room. A motor does the same, in two stages: the thin copper windings heat up within seconds, the heavy metal housing over many minutes. The heat flows out through a chain of thermal resistances (kelvin per watt: how many degrees warmer a part runs for each watt it must get rid of): winding to housing, then housing to the heat sink and the air. Each body also stores heat (a thermal capacitance). The simplest useful model has two nodes, a small fast winding and a large slow housing, the same structure motor makers publish as “winding” and “motor” thermal time constants:^[12]

Go deeper: the two-node heat equations

C_w dT_w/dt = P − (T_w − T_h)/R_wh C_h dT_h/dt = (T_w − T_h)/R_wh − (T_h − T_amb)/R_ha

T_w, T_h, T_amb: winding, housing and ambient temperature (°C). C: heat capacity (J/K). R: thermal resistance (K/W). The time constants are τ_w = R_whC_w (seconds) and τ_h = R_haC_h (minutes). dT/dt means “how fast the temperature changes”: heat in minus heat out, divided by the heat capacity.

In steady state the winding sits P × (R_wh + R_ha) above ambient. The rated torque is defined by that steady state: the torque whose heat brings the winding to its limit. RobStride’s manuals give a 145 °C winding limit (the protection threshold; the insulation itself survives 180 °C) at 25 °C ambient, a 120 K rise.^[2] That is why the rated torque depends on the heat sink: the manual’s RS04 makes 40 N·m bolted to a 345 × 345 mm aluminium plate but 35 N·m on a 220 × 200 mm plate. A bigger R_ha needs less heat for the same 120 K, and torque goes with the square root of heat: 35/40 = 0.875, so the smaller plate has about 1.31 times the thermal resistance.

Two more effects the playground includes. Above rated, the time to the limit depends on how far above: the manual’s RS04 table (rotating at 50 rpm) gives 108 s at 60 N·m and 15 s at 100 N·m. Holding still is worse than moving: in a stall, the current stops rotating between the three phases and one phase carries the peak continuously, so it makes √2 times the average heat. The manual’s stall-rated torque is 28.5 N·m for the RS04, close to 40/√2 = 28.3.^[2]

Driver modes: MIT mode

RobStride’s default “operation control” mode is the MIT mode, named after the drivers of MIT’s Mini Cheetah robot: every command carries a target position p, a target velocity v, stiffness k_p, damping k_d and a feed-forward torque τ_ff, and the driver computes, at its own fast rate,

τ_ref = k_p(p_set − p) + k_d(v_set − v) + τ_ff

and turns τ_ref into current.^[2] With k_p = k_d = 0 it is a pure torque motor; with k_p = 0 and v_set = 0 it is a damper; with k_p > 0 it is a spring around the target, which is what a walking policy uses (page 15). Each value is packed into 16 bits over a fixed range per model; for the RS04 the velocity range is ±15 rad/s and for the RS03 ±20 rad/s (robot/drivers/robstride.py, from the manuals), drawn in orange in the playground. Other modes (position, velocity, current) exist but MIT mode is the one that suits a learned policy.

Playground 2: winding temperature over a duty cycle

A duty cycle is the pattern of work and rest a joint goes through, such as “push hard for 40% of each step, then ease off”. Start here: press Static 18 cm step hold and then Marvin Bot today, and compare where the red line ends up.

How to read this picture: time runs left to right (minutes), temperature bottom to top. The red line is the copper winding, the blue line the housing, and the amber dashed line the 145 °C point where the driver protects itself and cuts torque. A red line that levels off below amber is fine; one that crosses it means the joint would stop mid-task. RMS (root mean square) torque, in the readouts, is the average that matters for heat, because it weights big torques more.

Presets

Marvin Bot today: knee climbing stairs Static 18 cm step hold Same, thigh-sized heat sink Knee while walking Hip yaw RS03 on one leg

Actuator and mount

Model RS04 (rated 40 N·m)RS03 (rated 20 N·m)RS10P (rated 14 N·m)RS00 (rated 5 N·m)

Heat path vs manual's plate

The continuous rating your mount achieves, as % of the manual's. RS04 on the manual's smaller 220 × 200 mm plate: 87.5% (35 of 40 N·m).

Ambient temperature

Torque profile

Profile ConstantWalking: 60% stance at full, 40% swing at 20%Stairs: 40% of each 3 s step at full, 60% at 30%Hold 30 s, rest 30 s

Torque at the high part

Holding still (stall: hottest phase √2 × heat)

Duration

Winding and housing temperature over time

–

RMS torque

–

RMS / rated in this mount

–

Winding at the end

–

Time to 145 °C

–

Steady-state winding

–

Rated torque in this mount

–

windinghousing145 °C protection limit

Model and assumptions

Two-node model above. Its constants are fitted so that (a) the rated torque gives exactly a 120 K steady rise at 25 °C ambient, including copper’s resistance rise, and (b) the times to the limit match the manuals’ overload tables (RS04 rotating at 50 rpm; RS03 at 100 rpm) within about ±40%. Fitted: RS04 R_wh = 10% of the total, τ_w = 2 s, τ_h = 250 s; RS03 20%, 20 s, 1000 s. RS06 and RS00 reuse the RS03 constants (their tables were not fitted: an assumption). Current follows the manual’s two points (rated and peak current), so saturation is included. The heat-path slider scales the housing-to-ambient resistance. The model starts cold, has no airflow from motion, and the stall factor is a simple √2; the manuals’ stall times are shorter than this model predicts, so treat stalls as optimistic.

Try this

1. Press Static 18 cm step hold: 47 N·m held continuously, 118% of rated. The winding reaches 145 °C after about 6 minutes. Press Same, thigh-sized heat sink (87%, a 35 N·m rating): the limit comes about a minute sooner, and the steady state it is heading for is far higher.
2. Press Marvin Bot today: an illustrative stair profile whose knee load averages about 74% of rated (RMS), the figure measured for the stair policy rough v5 on the earlier 1.29 m model. It warms up and levels off well below the limit: the duty cycle, not the peak, sets the temperature.
3. Press Knee while walking: an illustrative profile set to 70% of rated (RMS), the hottest joint (a knee) in the flat policy v9’s 1.0 m/s fast walk on the 1.41 m model (sim/isaaclab/README.md).
4. Tick Holding still on the static hold. The same torque without rotation heats one phase √2 times faster.
5. Set the RS04, constant profile, 120 N·m (peak). The model gives a few seconds: peak torque is for jumps, pushes and catching a fall, never for standing.

How Marvin Bot does it

In the repo

Four RobStride quasi-direct-drive models plus Feetech servos, specifications in description/actuators.yaml (from RobStride’s manuals^[1]) and placement in description/robot.yaml: RS04 at hip pitch, hip roll and knee (6), RS03 at hip yaw, both ankle motors, the waist and shoulder pitch (9 + 1 spare), RS10P at shoulder roll, shoulder yaw and elbow (6 + 1 spare), RS00 at wrist yaw (2), STS3250 servos at the neck and wrist pitch (4). All on 48 V, MIT mode over CAN at 500 Hz (legs) and 250 Hz (arms). The rule is: peak torques at most 70% of peak rating; continuous (RMS) torque within rating (docs/design/stairs.md).

Playground 3: why each model sits at each joint

Bars show the repo’s worst static demand for each joint as a percentage of the actuator’s continuous rating (filled) and peak (outline marker). Leg demands come from the static 18 cm step check (sim/mujoco/check_step_up.py, stance leg, current 1.41 m model); arm demands are the URDF’s holding torques (arm straight out, 0.5 kg in the hand), against the RS10P’s held-still rating where it holds a load. Change the robot’s mass, the heat path or the knee stage and see which joints cross 100%.

Presets

Marvin Bot today 35 kg target Real thigh heat sink With 1.5:1 knee stage

Robot mass (bare)

Leg demands scale with mass (about 1.2 N·m per kg at the knee in this pose; the repo rounds it to ~1.1).

Heat path vs manual's plate

Add the parked 1.5:1 knee stage

Knee: 60 N·m rated, 180 N·m peak, 14 rad/s, about +0.2 kg per leg.

demand ≤ 85% of rated85-100%over rateddemand as % of peak (rule: ≤ 70%)

Model	Peak N·m	Rated N·m	No-load rpm	Mass kg	Price	kg per 10 N·m peak	$ per N·m peak	$ per N·m rated	Joints
RS04	120	40	200	1.42	$255	0.12	2.13	6.4	hip pitch, hip roll, knee
RS03	60	20	195	0.90	$225	0.15	3.75	11.3	hip yaw, ankle ×2, waist, shoulder pitch
RS10P	42	14 (9.5 held)	125	0.46	$234	0.11	5.57	16.7	shoulder roll, shoulder yaw, elbow
RS00	14	5	315	0.31	$139	0.22	9.93	27.8	wrist yaw

Prices: Seeed Studio, October 2026; RS10P a reseller listing, November 2026 (bom/bom.xlsx). Torque and mass: description/actuators.yaml. Bigger actuators are cheaper and lighter per N·m, which is one reason the design uses the RS04 wherever the joint is heavily loaded, and why it stops at four sizes.

Reading the chart, the joint-by-joint story:

• Knee (RS04): the hardest working joint. In the static 18 cm step it holds 47 N·m: 39% of peak (fine) but 118% of continuous rating. The trained stair policy rough v5 loaded it only part of each step, 73-75% of rated RMS (measured on the earlier 1.29 m model); the current stair policy, rough v9, replayed on today’s 1.41 m model keeps the knees at 56-65% of rated (RMS over each stairs run, which includes the flat walk before and after). The real rating in a thigh-sized mount may be 30-35 N·m rather than 40, which would push the stair RMS to 85-100% (docs/design/stairs.md). The phase 2 bench test (bench/actuator/bench_robstride.py thermal --torque 47) measures it.
• Hip roll (RS04): 42 N·m on one leg, 105% of rated; about 60% RMS on stairs in simulation (rough v5, 1.29 m model).
• Hip pitch (RS04): lightly loaded in a static step (at most 12 N·m, in the swing-over-the-edge pose) but takes large dynamic peaks in walking and when standing up; the RS04 also keeps one model for the three big hip and knee joints.
• Hip yaw (RS03): 17.7 N·m in the static step, 89% of the RS03’s 20 N·m. It swings with the leg, so a lighter, smaller motor here also reduces leg inertia.
• Ankles (2 × RS03 through a linkage): 24 N·m pitch and 17.5 N·m roll needed against 32 and 25 N·m continuous available at the worst pose (tools/ankle_linkage.py; page 06). RS06 ankles were tried in v0.2 and rejected: 22 / 11 N·m continuous was not enough for slow stairs and one-foot standing (docs/decisions.md, 2026-10-26).
• Shoulders (RS03 pitch, RS10P roll and yaw) and elbow (RS10P): holding the arm straight out takes about 6.5 N·m at the shoulder pitch, 5.7 at the roll and 1.8 at the elbow (8.9, 8.0 and 3.2 N·m with 0.5 kg in the hand, from the URDF link masses). The RS10P’s 25:1 gear makes it light (0.46 kg) and strong for its size (42 N·m peak), but it holds only 9.5 N·m indefinitely while still (14 N·m while turning) and turns at most 13 rad/s: enough for the v1 hands’ 0.5 kg held out sideways, the 1 kg of requirement R20 for about two minutes, and every arm motion in the reference clips but a jog’s arm swing, which reaches its top speed. Shoulder pitch stays an RS03: 1 kg at full forward reach takes 11.3 N·m (docs/decisions.md, 2026-11-02).
• Waist yaw (RS03): not loaded by gravity; sized for the reaction of swinging arms and turning the torso. No static check in the repo.
• Wrist yaw (RS00) and wrist pitch / neck (STS3250 servos): light loads; the wrist pitch servo is at 75% of its rating with 0.5 kg at the fingertips (docs/requirements.md R20).

The simulator sees these actuators as peak torque and no-load speed per joint (tools/build_urdf.py), a straight-line DC-motor model, an armature (reflected inertia) placeholder per joint, and a 0-10 ms command delay. A heat penalty in training discourages holding more than about 0.9 times rated on average over one second (sim/isaaclab/README.md).

Trade-offs and alternatives

Option	For	Against	Status
RobStride RS04 / RS03 / RS10P / RS00 (QDD)	most torque per dollar of the QDD actuators compared; public manuals and STEP models; Python, Rust and ROS 2 drivers	ratings depend on the heat sink; some figures differ between README and manual	chosen (docs/decisions.md, 2026-10-25)
Other QDD brands (CubeMars, Damiao, Encos, MyActuator)	second sources; the Damiao DM-J10010L matches the RS04 class	2-3× the price per N·m or fewer verified specs	DM-J10010L kept as second source
Self-built cycloidal actuators	cheap, open (as in Berkeley Humanoid Lite[9])	more work, lower torque, wear	rejected
Parked 1.5:1 knee stage (belt or linkage after the RS04)	60 N·m continuous, 180 peak	~14 rad/s top speed, ~0.2 kg, some backlash, thigh redesign	parked until the bench test; also needed if legs grow longer (127% of rated at +90 mm legs)
Better heat path (thigh plate as heat sink, paste, small fan)	+10-20% continuous, cheap	to be measured	first option to try
RS06 at shoulder yaw and elbow, RS03 at shoulder roll (until 2026-11-02)	more speed (50 and 20 rad/s) and peak torque at the shoulder roll	1.52 kg heavier	replaced by RS10P
RS10P at shoulder pitch and waist yaw too	saves ~0.9 and 0.44 kg	shoulder pitch over its held rating with 1 kg at full reach; the waist’s bearing and hub are built round the RS03	checked and declined (2026-11-02)

Where sources and the repo disagree

• RS03 no-load speed and mass: RobStride’s specification README says 195 rpm and 900 g; the RS03 manual (CAN 2.0, revision 261013) says 200 rpm and 880 g. actuators.yaml uses the README’s 195 rpm and 0.90 kg and records both.^[1][3]
• RS03 rated torque: README 21 N·m, manual 20 N·m (215 × 220 mm plate); the repo uses 20.
• Voltage range: the README gives 15-60 V for the RS03 and RS04; their manuals say 24-60 V, which the repo uses (15-60 V for the RS10P, matching its manual).
• RS10P torque constant and speed: the RS10P manual prints K_t 7.73 N·m/A and its specification sheet 3.73; the repo uses 3.73. Its manual’s 48 V T-N curve runs to about 315 rpm, against 125 rpm no-load speed in the same manual and the specification sheet’s table (14 N·m at 100 rpm, 3 at 120); the repo uses 125 rpm.^[15]
• Torque-speed model: the manuals’ T-N curves keep near-peak torque up to about half the no-load speed; the simulation’s straight-line DC-motor model (saturation torque = peak) is much more conservative at mid speeds. Not an error, but a modelling choice to revisit with bench data.
• Ankle demand: docs/design/stairs.md says the ankle needs “~22 and ~14 N·m” (pitch, roll); running check_step_up.py on today’s 1.44 m model gives 24.4 and 17.5 N·m, still within the 32 / 25 N·m available.

Decisions you can make here

• Whether the knee needs the 1.5:1 stage: compare the stair RMS (Playground 2) against the rating you expect in the real thigh mount. Trade: 50% more continuous knee torque, but a third less top speed, about +0.2 kg per leg and a thigh redesign.
• How much heat path to design in: every percent of rating gained by a better plate or a small fan is torque for little money, at the cost of some mass and design time.
• Whether to give the simulator the manual’s torque-speed (T-N) curve instead of the straight line: faster, less cautious policies, but more risk of asking for torque the real motor lacks if the curve is optimistic.
• Gear ratio for a new joint: torque goes up by N, but the foot-strike “mass” goes up by N² and the joint stops giving way when bumped.
• Battery voltage window: letting the battery run lower gives more runtime per charge but costs top speed (not torque) near the end.

Your decision noteCopy

Built from your settings above. Paste it into an issue, a design discussion or an AI assistant session.

Where this leads. This page treated each motor on its own. Page 05 works out where the leg loads come from (the robot’s weight times lever arms) and why stairs are the hardest case, and page 06 shows how two of these motors share the ankle’s work.

Further reading

• Wensing et al., Proprioceptive actuator design in the MIT Cheetah (2017): the clearest explanation of why low gear ratios suit legged robots, with the impact and bandwidth analysis.
• Lynch and Park, Modern Robotics, section 8.9: motors, gearing, torque-speed curves and reflected inertia in textbook form.
• RobStride Product_Information: the manuals with T-N curves, overload tables and the CAN protocol, worth reading before the bench test.
• Chi et al., Berkeley Humanoid Lite (2025): an open humanoid with self-built cycloidal actuators, including their backlash and wear testing.
• Isaac Lab actuator models: what the simulator does with your torque and speed numbers.

Sources

1. RobStride, Product_Information repository, specification README — used for: peak, rated, no-load speed, ratio, mass and voltage per model; README-versus-manual differences.
2. RobStride, RS04 User Manual (CAN 2.0, rev. 260713) — used for: K_t 2.1 N·m/A, rated/peak currents, back-EMF, T-N curve, 345 / 220 mm plate ratings, 145 °C limit, overload and stall tables, operation-control (MIT) formula, field-weakening parameter.
3. RobStride, RS03 User Manual (CAN 2.0, rev. 261013) — used for: RS03 K_t, 17 V/krpm back-EMF, 200 rpm and 880 g, T-N curve, 215 × 220 mm plate, overload table.
4. RobStride, RS06 User Manual (CAN 2.0, rev. 260713) — used for: RS06 K_t, currents, back-EMF, T-N curve, test plate (the arm motor until 2026-11-02).
5. RobStride, RS00 User Manual (CAN 2.0, rev. 260713) — used for: RS00 K_t, currents, back-EMF, T-N curve.
6. P. M. Wensing, A. Wang, S. Seok, D. Otten, J. Lang, S. Kim, “Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots,” IEEE T-RO 33(3), 2017 — used for: proprioceptive actuation, reflected inertia and impact mitigation, backdrivability.
7. S. Seok et al., “Design Principles for Energy-Efficient Legged Locomotion and Implementation on the MIT Cheetah Robot,” IEEE/ASME T-Mech, 2015 — used for: large-gap motors with a single low-ratio stage.
8. B. Katz, J. Di Carlo, S. Kim, “Mini Cheetah: A Platform for Pushing the Limits of Dynamic Quadruped Control,” ICRA 2019 — used for: low-cost modular QDD actuators and the MIT-mode driver lineage.
9. Y. Chi et al., “Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot,” RSS 2025 (arXiv 2504.17249) — used for: cycloidal gearboxes chosen for printed parts; directly driven joints.
10. K. M. Lynch, F. C. Park, Modern Robotics, Cambridge University Press, 2017, §8.9 — used for: gearing, reflected (apparent) inertia N²·J, harmonic drives.
11. Isaac Lab documentation, Actuators — used for: the DC-motor torque-speed clipping model.
12. maxon, catalogue motor data page (thermal data block) — used for: the convention of two thermal resistances and two time constants (winding, motor).
13. Texas Instruments, TIDUF56 design guide (field weakening and MTPA) — used for: field weakening extends operation above base speed.
14. NBS Handbook 100, Copper Wire Tables (1966) — used for: copper’s temperature coefficient of resistance, 0.00393 per K at 20 °C.
15. RobStride, RS10P User Manual (CAN 2.0, rev. 260907) and RS series specification sheet (2026-10-09) — used for: RS10P 42 / 14 N·m (9.5 N·m stalled), 125 rpm, 25:1, 0.46 kg, K_t, currents, back-EMF, T-N table, overload and stall tables, output-side inertia 0.00625 kg·m², mounting holes, MIT ranges.