WorksComputer Science
BiPathCPG: A Novel Load-Sharing Design Architecture for Compound Planetary Gearboxes
Abstract
Quasi-Direct Drive (QDD) actuators are widely used in legged robots due to their low reflected inertia, high backdrivability, and high torque density, commonly employing planetary or cycloidal gearboxes. Single-stage planetary gearboxes offer high efficiency but limited torque density; multi-stage designs improve torque density at the cost of efficiency; and cycloidal gearboxes provide high stiffness but suffer from torque ripple. Compound Planetary Gearboxes (CPG) offer a compromise, but conventional designs result in increased actuator width, mass, and reduced torsional stiffness in 3D-printed actuators. This paper proposes the BiPath Compound Planetary Gearbox (BiPathCPG), a novel architecture that introduces an idle second-stage sun gear and load-sharing second-stage planets within the otherwise unused space of a conventional CPG to simultaneously improve stiffness, reduce actuator width, and reduce mass. An optimization framework computes BiPathCPG parameters for a given motor and target gear ratio. The architecture is validated through a 3D-printed proof-of-concept actuator design, benchmarked against a conventional CPG using custom-built test rigs. The BiPathCPG achieves a torsional stiffness of 1394 Nm/rad, a 1.70x improvement over the conventional CPG, with a 32% reduction in backlash, 25 g lower mass, and approximately 6.4 mm reduction in axial width, while achieving a peak stall torque of 36.49 Nm.
Introduction
Industrial demand for robots, including humanoids and quadrupeds, is increasing. These robots require efficient, high-performance joint actuators, often with metal gearboxes [1], [2]. At the same time, growing interest in cost-effective legged robot research platforms [3], [4] has increased interest in actuators with 3D-printed gearboxes [5], [6], [7]. In both scenarios, one of the most widely adopted solutions is the Quasi-Direct Drive (QDD) actuator [8], [9], [10], [11], [12], which pairs a high torque-density motor with a low gear ratio, enabling low reflected inertia and high backdrivability.
In metal-gearbox QDD actuators, one of the most commonly used transmission architectures is the single-stage planetary gearbox (SSPG) [13], [11], [8], [9]. SSPGs are efficient and compact, but their torque density is limited by achievable gear ratios and peak torque capacity. Alternative planetary architectures, such as the Wolfrom planetary gearbox (WPG, 3K) [14], [15] and double-stage planetary gearbox (DSPG) [16], provide higher torque density through larger gear ratios, but WPGs typically exhibit lower efficiency and DSPGs suffer reduced efficiency from multi-stage power losses.
For QDD actuators with 3D-printed gearboxes, planetary gearboxes are often at a disadvantage compared with cycloidal gearboxes [17], [18], [6], which are typically more axially compact [19], [20], [7]. Axially compact cycloidal–planetary hybrid gearboxes have also been developed [21], [22]. Cycloidal gearboxes achieve higher gear ratios within a given space, while the lower strength of 3D-printed materials requires wider planetary gears and larger dimensions, increasing actuator width and mass. 3D-printed planetary actuators also exhibit lower stiffness than 3D-printed cycloidal actuators [3], [23], [5]. Cycloidal gearboxes, however, can exhibit output torque ripple [24], [1]. While we recognize the advantages of cycloidal architectures in axial compactness, this work focuses on addressing the limitations of 3D-printed planetary gearboxes.
Among planetary architectures, compound planetary gearboxes (CPGs) offer an effective compromise between efficient but low-torque-density SSPGs and high-torque-density but less efficient WPGs and DSPGs [25], [26]. CPGs (Figure 2(a)–(Figure 2c)) achieve higher gear ratios than SSPGs at comparable efficiency [23]. However, the smaller second-stage planet diameters require relatively wide second-stage planet and ring gears, increasing actuator mass in both metal and 3D-printed CPGs. 3D-printed CPG actuators also exhibit lower stiffness than 3D-printed cycloidal gearbox

Figure 2. Gearbox architectures: (a)–(c) CPG and (d)–(f) BiPathCPG, shown as front view, line diagram and isometric view, respectively. : sun, : first-/second-stage planets, : ring, : carrier. and denote the idle second-stage sun and planet gears of the BiPathCPG.
This work was supported by the DRDO Industry Academia Raman Centre of Excellence, IISc, under grant DFTM/02/3125/M/01/AIR-01.
We thank Divyansh Raghadale for his help with the experiments. actuators [3], [23], [5].
To address the large width and mass of compound planetary actuators (CPG), as well as the lower stiffness observed in 3D-printed CPG actuators, we propose a new design called the BiPath Compound Planetary Gearbox (BiPathCPG). This proof-of-concept architecture for compound planetary gearboxes employs idle load-sharing gears positioned within the unused space of a traditional CPG configuration, achieving stiffness and backlash characteristics comparable to cycloidal gearboxes while maintaining high efficiency. Our main contributions are:
We propose a novel compound planetary gearbox architecture, the BiPathCPG, which improves stiffness and reduces actuator width and mass.
We experimentally validate the design through a 3D-printed proof-of-concept prototype and compare it with a traditional compound planetary gearbox.
We develop a Python-based framework that computes the BiPathCPG design variables for a given motor and desired gear ratio. [Code-link]
Background
Compound Planetary Gearbox (CPG) Architecture
The compound planetary gearbox (CPG) is a planetary architecture that uses compound planets: two gears of different diameters rigidly connected to rotate at the same angular velocity. In the most torque-dense configuration, the sun drives the larger planet, the smaller planet meshes with a stationary ring, and the output is taken from the carrier. Figure 2(a)–(Figure 2c) shows a 2D front view and line diagram of this configuration.
The gear reduction ratio is given by
where is the number of teeth on the sun gear, and those on the larger and smaller planet gears, and and the modules of the sun–planet and planet–ring meshes, taken equal () for simplicity. The overall efficiency is
where and are the sun–planet and planet–ring mesh efficiencies, calculated using the formulation in [27].
Gear Tooth Strength Modeling
Gear width is an important gearbox design parameter. It depends on the operating load. We determine it using the modified Lewis gear strength equation [28], as follows:
To calculate the face-width of the gear we can rearrange the (3) and add a Factor of Safety (), as follows:
where is the bending strength of the gear material, is the tooth force applied to the gear tooth, is the width (face-width) of the gear, is the module of the gear, is the factor from contact ratio, and is the tooth form factor, calculated using the figure given in [28], [Drivetrain Hub]. Here, .
Problem Formulation
Increased Mass and width of the CPG Actuator
Static Force Modeling of the Compound Planetary Gearbox (CPG)
1) Static Force Modeling of the Compound Planetary Gearbox (CPG): Gear face width is set by the tooth forces, which follow from a steady-state force balance: the motor applies to the sun gear while the load at the carrier holds every gear at constant angular velocity. From the free-body diagram in Figure 3, the equilibrium equations are

Figure 3. Free body diagrams of (a) CPG and (b) BiPathCPG; (c)–(d) gear tooth tangential contact forces.
where:
, , : tangential contact forces at the sun–planet 1, planet 2–ring, and planet–carrier interfaces
, , : radii of the sun, planet 1 (larger) and planet 2 (smaller) gears
: number of planet gears
, : motor input torque and output load torque
Under steady-state operation, with all gears at constant angular velocity, the load torque satisfies
where denotes the gear ratio. For simplicity, transmission losses are neglected; therefore, efficiency is not included in the output torque calculation. To determine the maximum tangential forces acting on the gear teeth, the input torque is taken as the peak torque rating of the motor used in the actuator.
Solving these equations gives
From (7), is times the magnitude of . Since , . Moreover, increases as the number of teeth decreases [28]. According to (3), these factors require a significantly larger face width for the second-stage planet gear. Since the ring gear must match the face width of the meshing planet gear, its width also increases. This, in turn, increases the overall actuator width and mass, as components such as the carrier and outer housing must be sized accordingly.
Stiffness Modeling of CPG Gearbox
To estimate the mesh-induced stiffness of the CPG, we adopt a simplified lumped-parameter formulation similar to [29], in which each gear mesh is represented by a linear spring acting along the tangential direction. Only the compliance of the gear meshes is considered, while the deformation of the gears, planet shafts, carrier, bearings, and housing is neglected. The total elastic strain energy stored in the CPG meshes is expressed as
where and denote the linear mesh stiffnesses of the sun–planet-1 and planet-2–ring meshes, respectively [30]. Using the mesh forces obtained in (7), the input angular deflection is obtained from Castigliano’s theorem [31] as
The resulting input-referred torsional stiffness is
The corresponding output-referred stiffness is
Since in a CPG, the contribution of the second-stage mesh compliance is amplified by the factor . Therefore, for comparable mesh stiffnesses, the CPG exhibits lower torsional stiffness than a single-stage planetary gearbox, for which .
BiPathCPG Gearbox
To address the issues discussed in the previous section, we introduce the BiPathCPG. Idle planet gears are added to the second stage, Figure 2(d)–(Figure 2f) (pink), equal in number to the planets of a conventional CPG and identical to the second-stage planets in diameter and tooth count. An idle second-stage sun gear is also added, Figure 2(d) (transparent); it is not rigidly attached to the first-stage sun but rotates about the gearbox centre on a bearing and meshes with all second-stage planets, redundant ones included. The idle planets open a second load path in the second stage, reducing the force carried by each primary ring–planet mesh and so allowing a smaller stage-2 face width, and with it a narrower and lighter actuator.
Gear-Width Analysis of the BiPathCPG
The static force equilibrium of the BiPathCPG gearbox can be written as:
where:
, , : contact forces at the sun–planet 1, planet 2–ring, and planet–carrier interfaces
, , , : contact forces involving the idle sun and idle planet gears (idle sun–planet 2, planet 2–idle planet, idle planet–ring, idle planet–carrier)
, , , , , : radii of the sun, planet 1, planet 2, idle sun, idle planet, and ring gears
, : motor input torque and output load torque
Equation (12) can be written compactly as
where , , and . The matrix is rank deficient, implying that the system admits infinitely many solutions. This occurs because the BiPathCPG gearbox contains kinematically redundant gear interactions.
A similar case occurs in a four-legged table with an uneven load: equilibrium alone cannot uniquely determine each leg’s force without considering stiffness or contact model. Likewise, the BiPathCPG gearbox needs additional physical assumptions to determine a unique force distribution.
Load-sharing assumption
To resolve this indeterminacy, we introduce the load-sharing relation
where is the force-sharing factor between the planet–ring and idle-planet–ring contacts. Although the idle planets are kinematically redundant, is generally nonzero: gear-body and tooth compliance, bearing friction, manufacturing tolerances and profile deviations distribute load over all stage-2 meshes. Thus indicates that the idle mesh carries less load than the primary planet mesh, more, and the limiting case recovers the conventional CPG, in which the entire ring reaction is carried by the planet meshes alone.
Applying (13) reduces (12) to
Solving it further, we get:
The total ring reaction is independent of and equals that of a conventional CPG: the idle planets just redistribute it over meshes. Since the face width of each mesh is sized by its own contact load, this redistribution makes the second stage axially thinner.
FEA estimation of the force-sharing factor
A simplified model retaining one set of meshing gears (Figure 4a) was analysed in SolidWorks Simulation: the compound planets –, the idle planets and the stage-2 idle sun rotate freely, the stage-1 sun and the ring are fixed, 15 N·m is applied to the carrier , and all meshes start at zero backlash with frictionless contact. With equal planet and idle-planet face widths of 8.2 mm the ring-mesh forces give ; rescaling the face widths in proportion to those loads (Figure 4b) gives , and a third study with the carrier fixed and 3 N·m applied to the sun reproduces the same ratio. Because the model is frictionless this is a lower bound: the idle path closes through the freely rotating , and bearing friction there raises the torque carried along the idle branch. We therefore adopt .

Figure 4. FEA study of stage-2 load sharing.
With the planet–ring contact force is 62.5% of the value in a conventional CPG of identical gear geometry, the idle mesh carrying the remaining 37.5%. The stage-2 planet face width, and hence the axial width of the ring and of the outer shell, is reduced accordingly, while the idle planet requires only 0.6 times the planet face width.
Stiffness modeling of BiPathCPG Gearbox
The lumped mesh-stiffness formulation of Section III-B is applied to the BiPathCPG, with the strain energy now summed over the , , , and meshes and the mesh forces shown in (15). Writing for the total compliance of the idle path, , and defining
the resulting input-referred torsional stiffness is
with the output-referred stiffness following as .
Stiffness Comparison of CPG and BiPathCPG
To obtain a simplified analytical comparison between the two architectures, all gear meshes are assumed to have the same linear mesh stiffness, , and is defined. This equal-stiffness assumption is a simplification, since the actual mesh stiffness depends on the individual gear geometry and contact conditions. However, it provides a useful first-order comparison of the relative stiffness of the CPG and BiPathCPG architectures. The input-referred stiffness of the CPG then becomes
while, for the BiPathCPG with ,
The stiffness improvement can therefore be expressed as
Since in a CPG, , for which the above ratio is greater than unity. Thus, under this simplified equal-mesh-stiffness assumption, the BiPathCPG exhibits higher mesh-induced transmission stiffness than the conventional CPG.
For the fabricated design of Table 1, where and share a common module, , and the above ratio evaluates to 1.18, so the BiPathCPG is predicted to be 1.18× stiffer than the conventional CPG.
| Parameter | CPG | BiPathCPG |
| 14.00 | 14.00 | |
| (mm) | 0.6 | 0.6 |
| 18/66/33/117/– | 18/66/33/117/51 | |
| 3 | 3 | |
| (mm) | 11.2/11.2/16.4/16.4/–/– | 11.2/11.2/10.2/10.2/10.2/8.2† |
| (mm) | 87.10 | 80.94 |
| (kg) | 0.815 | 0.801 |
| (N·m) (Pred.) | 36.96 | 36.96 |
| (Theoretical) | 0.938 | 0.938 |
| (N·m/kg) | 45.35 | 46.14 |
Table 1. Optimized Parameters, :1, MAD M6C12
Optimization and CAD Design
The many interdependent design parameters of CPG and BiPathCPG architectures, together with the additional geometric and meshing constraints of the BiPathCPG, make manual parameter selection difficult and error-prone, motivating a computational design framework. We therefore extend the COMPAct framework [23] to the BiPathCPG. Following the same work, the optimizer minimizes the cost function:
where , , and are the actuator mass, mechanical efficiency, axial width and gear ratio, and , , , are weighting coefficients. and follow Eqs. (2) and (1), and and the mass and geometric model of [23]. We set kg, and mm, balancing the three terms over the actuator sizes considered; when a target ratio is specified () and 0 otherwise.
The major changes and results of this framework are explained below.
Constraint Modifications
Constraint Modifications: In the conventional CPG the second-stage planet count equals that of the first stage, , whereas the BiPathCPG doubles it, . This change, along with the addition of a stage-2 sun gear, propagates into three classes of constraints.
a) Geometric Constraints. The first geometric relation is identical in both configurations:
The BiPathCPG additionally enforces the stage-2 geometric constraint:
a condition absent in the conventional CPG due to the lack of a stage-2 sun gear.
b) Meshing Constraints. The conventional CPG requires for proper meshing, while the BiPathCPG imposes the stricter conditions
where the second condition is the standard sun–planet–ring meshing compatibility applied to stage 2 with equally spaced planets.
c) No Interference Constraints. The stage-1 planet-to-carrier clearance constraint is unchanged:
The BiPathCPG additionally enforces the corresponding stage-2 clearance constraint:
Here, and denote the pitch radii of the sun and planet gears in stage , mm is the carrier extrusion radius, and mm is the minimum clearance to avoid interference. Note: The no-interference Eq. (26) and meshing Eq. (24) constraints can lead to radially larger designs.
Optimization Results and CAD Design
Optimization Results and CAD Design: Two simulation experiments were conducted for every motor in our dataset, across gear ratios from 4:1 to 20:1 in unit increments. The first used 3D-printed material properties (PLA: kg/m, MPa); the second assigned steel to gears ( kg/m, MPa) and aluminum to remaining components ( kg/m). Both experiments used identical constraints: , mm, and . The full sweep is shown in the supplementary video; from these results we selected the MAD M6C12 motor at :1 for fabrication.
The fabricated design is shown in Figure 5, with its parameters in Table I and the assembled actuators in Figure 1. The predicted axial widths (87.1, 80.9 mm) and masses (0.815, 0.801 kg) for the CPG and BiPathCPG agree closely with measured values (85.3, 78.9 mm) and (0.820, 0.795 kg), validating the geometry and mass model. The fabricated BiPathCPG is 25 g lighter with a 6.4 mm reduction in axial width.

Figure 1. Actuators (3D-printed): Compound Planetary Gearbox (CPG) (left, 820 g) and BiPathCPG (right, 795 g).

Figure 5. BiPathCPG: (a) exploded view, (b) sectional view.
Results
Based on the optimisation results, a gear ratio of :1 (Table I) with the MAD M6C12 150KV motor was selected for fabrication. Both the CPG and BiPathCPG were fully 3D-printed in PLA at 100% infill using a Bambu Lab X1C printer. All experiments were conducted using a custom, in-house developed BLDC motor driver.
Efficiency Comparison
Power efficiency was measured using a custom dynamometer that couples the test and load actuators output-to-output. A static reaction torque sensor (STARK EMBASYS: SE-TS-F) is mounted between the test actuator housing and a rigid ground fixture. The load actuator maintains constant angular velocity while the test actuator runs in torque-control mode; output power is reaction torque times encoder-derived velocity (scaled by gear ratio), and input power is commanded motor torque times motor velocity. Mechanical efficiency is the output-input ratio. The setup follows prior work [23, 3], with a static torque sensor replacing the load cell. The results of this experiment are shown in Figure 6(a) and (b). The conventional CPG achieves a mechanical efficiency of approximately at higher torques, whereas the BiPathCPG achieves approximately . The lower efficiency of the BiPathCPG can be attributed to the presence of additional idle planet gears in the second stage, which increases friction compared with the conventional CPG. The theoretically modeled efficiency values reported in Table I are the same for both gearboxes because the efficiency model used in this work [27] does not account for the number of planet gears. Nevertheless, the difference in mechanical efficiency between the two gearbox architectures remains relatively small.
Figure 6. Bench comparison of the CPG and BiPathCPG at . (a), (b) Mechanical efficiency versus commanded torque at 20–120 rpm; the shaded band marks the saturated region above 12 N·m and the dash-dot line its mean. (c), (d) Output deflection under a N·m fixed-output load cycle. (e), (f) Static load-cell torque during the backdriving ramp, with circles marking motion onset. (g) Output deflection versus torque with forward and backward linear fits. (h) Load-cell torque for a 40 N·m command.
Backlash and Transmission Stiffness
Both quantities were measured on the same fixed-output rig following [5], with the output rigidly constrained by a steel plate fixed to the ground. The applied torque was ramped to the target over 7 s, held for 3 s, and returned to zero, then repeated in the negative direction, giving a 33 s load cycle. Each cycle was repeated three times at each of three output orientations spaced apart, giving nine measurements per gearbox. Backlash was measured at N·m (about 0.14 N·m at the motor), low enough that gearbox compliance contributes negligibly and the deflection is backlash alone; stiffness was measured at N·m. Deflection was read from the 16-bit motor-side encoder, resolving at the motor side.
For each measurement, backlash was computed as the difference between the maximum and minimum output-referred angular deflection over the load cycle, and the reported value is the average of the nine measurements (Figure 6(c) and (d)). The measured backlash is for the BiPathCPG and for the CPG, a 32% reduction. This primarily results from the increased number of simultaneously engaged gear meshes in the BiPathCPG.
Stiffness was estimated by fitting to the encoder deflection separately for each direction of loading, taking as the mean of the forward and backward fits. The conventional CPG yielded N·m/rad (801 forward, 843 backward), while the BiPathCPG reached N·m/rad (1460 forward, 1328 backward), a 1.70× increase (Figure 6(g)), above the 1.18× predicted in Section IV-B. The difference can be attributed to the fact that the model assumes equal tangential mesh springs, no shaft or gearbox-body compliance, and omits an effect that acts only on the BiPathCPG: under ring load the stage-2 planets deflect radially inward in the CPG, whereas the idle sun braces them from within. This likely accounts for much of the difference. The BiPathCPG’s stiffness exceeds the Berkeley Humanoid Lite and COMPAct actuators by more than 4×, and reaches 95% of the PA6-CF non-pinwheel cycloidal actuator [19] (Table II), despite relying on spur-gear meshing rather than the inherently stiffer rolling contact of cycloidal drives.
D×L is the actuator diameter × axial length; the value in brackets, Lg, is the axial length of the gearbox alone. In the CPG and BiPathCPG of this work the motor–gearbox coupling sits outside the gearbox and is therefore counted in L but not in Lg; in designs with an internal coupling the coupling is already contained in Lg. ∗Predicted peak torque is the manufacturer-specified motor peak torque multiplied by the gearbox ratio. †The referenced Bezzini papers report gearbox mass only; a motor mass of 638 g has been added to obtain the actuator mass. ‡The referenced Bezzini papers report gearbox dimensions only; a motor axial width of 52.3 mm has been added to obtain the actuator height to obtain the actuator axial length. For CPG, PLACyC and CP30 the reported gearbox diameter is smaller than the φ90 mm motor, so the actuator diameter is set by the motor rather than the gearbox. §Stiffness and backlash reported for the conic-disk anti-backlash variant at 100% preload. ¶Backlash reporting is not uniform across sources: mean ± standard deviation where available, and worst-case or typical play otherwise.
Stall Peak Torque Test
We evaluate stall peak torque with a custom rig: the actuator is rigidly mounted to a fixed base, and a 30 cm radial arm on the output shaft presses against a load cell at its distal end. Output torque is the product of the measured contact force and the arm length. During the experiment, the commanded torque is increased linearly from zero to a predefined maximum value, held at the peak for s, and then decreased linearly back to zero. This procedure allows us to estimate the stall peak torque of the actuator. Figure 6(h) shows the measured torque profiles. The conventional CPG and the BiPathCPG reach peak torques of and N·m respectively, averaged over three trials. The two architectures are therefore equivalent in peak torque capacity, with the BiPathCPG matching the conventional CPG despite its lower mass and axial width.
Backdrivability Test
Backdrivability was measured on the same setup used for the efficiency tests, with a load motor mounted to a static torque cell and coupled to the output of the actuator under test. Before each trial, the actuator was rotated in the loading direction so that the gear teeth were already in contact and any backlash was taken up. The load motor then ramped the output torque from 0 to 1.0 N·m over 25 s, while load position and torque-cell readings were logged. The test was repeated three times clockwise and three times counter-clockwise for each gearbox. The actuator was considered back-driven at the first instant where the load speed, averaged over 0.5 s, exceeded 0.5°/s and remained above this threshold for 10 consecutive samples; the back-drive torque is the load-cell torque averaged over the 0.5 s window at that instant. The results are shown in Figure 6(e) and (f). The mean back-drive torque is 0.59 N·m for the BiPathCPG (0.65 N·m CW, 0.52 N·m CCW) and 0.52 N·m for the CPG (0.60 N·m CW, 0.43 N·m CCW). As expected from the efficiency tests, the BiPathCPG is slightly harder to back-drive than the CPG.
Steady-State Operation and Failure Tests
Durability was evaluated using steady-state and failure cyclic tests. In both, the actuator oscillated between +45° and −90° while driving a 1 kg payload. In the steady-state test, the payload was mounted on a 30 cm extrusion and oscillated at 0.3 Hz, producing a measured peak torque of approximately 8 N·m per cycle. Both gearboxes operated for 12 hours without failure and showed negligible wear. In the failure test, the payload was mounted on a 75 cm extrusion and oscillated at 0.34 Hz. The quasi-static gravitational torque was approximately 11 N·m, while arm inertia produced additional spikes at direction reversals, raising the measured peak torque to approximately 28 N·m. Results of both tests and representative torque-versus-time profiles are provided in the supplementary video. Under this loading, the BiPathCPG operated for 45 minutes before the stage-1 sun gear stripped, with negligible wear in the remaining gears. The conventional CPG failed similarly after 37 minutes. This is likely because both designs share an identical first stage (Table I), with the smallest and most heavily loaded stage-1 sun determining the torque-limiting failure mode.
Actuator Comparison
Table II positions the BiPathCPG among axially compact 3D-printed gearboxes. Its transmission stiffness is the second highest listed and the highest of any all-PLA design. Its backlash is lower than that of every pure cycloidal entry; only designs operating at substantially higher gear ratios achieve lower values. In axial width the cycloidal designs are the clear winners at comparable gear ratios, and they remain competitive at much higher ones, reaching reductions of 30:1 and above within a similar axial envelope. For mass the fairer comparison is torque density, which is similar to that of pure cycloidal gearboxes at comparable gear ratios; because cycloidal designs reach high reductions in very compact volumes, however, matching their torque density at high gear ratios is not attainable with the BiPathCPG. Against its own CPG baseline the BiPathCPG nevertheless improves stiffness, backlash, mass and axial width, as reported above.
| Actuator | Motor (Peak Torque) | Gearbox-type (gear material) | Ratio | Stiffness (N·m/rad) | Backlash¶ (°) | Mass (g) | D×L (Lg) (mm) | Exp. Peak Torque (N·m) | Pred. Peak Torque∗ (N·m) |
| Roozing Cycloidal [20] | MN8012 (3.82 N·m) | Cycloidal (PLA) | 11 | 633 | 0.35 | 809 | 104 × 58(23) | > 40 | 42.02 |
| Roozing (Non-Pinwheel) [19] | MN8012 (3.82 N·m) | Cycloidal (PA6-CF) | 11 | 1468 | 0.25 | 797 | 104 × 58(23) | 36.4 | 42.02 |
| Roozing Anti-backlash [5] | RO100 (12 N·m) | Cycloidal (PA6-CF) | 13 | 300§ | 0.647§ | 1350 | 125 × 62(25) | 156 | 156 |
| Berkeley Humanoid Lite [3] | M6C12 (2.64 N·m) | Cycloidal (PLA) | 15 | 319 | 1.312 ± 0.241 | — | 88 × 66(35) | — | 39.60 |
| CY30 [7] | Maxon EC90 Flat (5.03 N·m) | Cycloidal (PLA) | 30 | 641.4 | 1.85 ± 0.39 | 942† | 92 × 85.8‡(33.5) | — | 150.90 |
| COMPAct SSPG [23] | MN8014 (4.97 N·m) | Planetary (single stage) (PLA) | 7.2 | 242.7 | 0.6 | 838 | 102 × 82(47) | — | 35.78 |
| EMBIR SSPG [6] | RI50 (1.67 N·m) | Planetary (single stage) (PLA) | 7.5 | — | — | 482 | 90 × 70(—) | 7.5(20s)/19.1(2s) | 12.53 |
| EMBIR WPG [6] | RI50 (1.67 N·m) | Planetary (Wolfrom/3K) (PLA) | 15.04 | — | — | 536 | 90 × 90(—) | 15.0(20s)/38.2(2s) | 25.12 |
| CO30 [7] | Maxon EC90 Flat (5.03 N·m) | Planetary (Wolfrom/3K) (PLA) | 30 | 723.2 | 0.53 ± 0.21 | 1002† | 100 × 105.9‡(53.6) | — | 150.90 |
| CFB [15] | Maxon EC90 Flat (5.03 N·m) | Planetary (Wolfrom/3K) (PLA) | 30 | 971.5 | 0.114 ± 0.076 | 899† | 90 × 100.7‡(48.4) | — | 150.90 |
| CP30 [7] | Maxon EC90 Flat (5.03 N·m) | Cycloidal-Planetary (PLA) | 30 | 573.6 | 1.10 | 910† | 78 × 104.3‡(52.0) | — | 150.90 |
| PLACYc [22] | Maxon EC90 Flat (5.03 N·m) | Planetary-Cycloidal (PLA+N12CF) | 45 | 918.1 | 0.066 ± 0.029 | 792† | 70.5 × 82.3‡(30.0) | — | 226.35 |
| CYPLO [21] | Maxon EC90 Flat (5.03 N·m) | Cycloidal-Planetary (PLA+N12CF) | 52.5 | 859.6 | 0.175 | 909† | 87 × 82.75‡(30.45) | — | 264.08 |
| CPG (Ours: Baseline) | M6C12 (2.64 N·m) | CPG (PLA) | 14 | 822 | 0.311 ± 0.004 | 820 | 104.7 × 85.3(48.7) | 36.22 | 36.96 |
| BiPathCPG (Ours: Proposed) | M6C12 (2.64 N·m) | BiPathCPG (PLA) | 14 | 1394 | 0.213 ± 0.014 | 795 | 104.7 × 78.9(42.3) | 36.49 | 36.96 |
Table II. Actuator Performance Comparison
Conclusion
This work presents the BiPathCPG, a novel architecture that improves conventional compound planetary actuators, particularly 3D-printed ones, by adding idle load-sharing planets and a sun gear within the otherwise unused volume of a CPG. This yields simultaneous gains in torsional stiffness, width, and mass, validated by fabricating and testing optimizer-selected BiPathCPG and CPG designs on custom-built rigs. Limitations include lower efficiency and slightly harder backdriving from the added idle meshes, and increased assembly complexity due to the additional idle gears. Under dynamic loading peaking at , both gearboxes stripped the stage-1 sun within an hour of cycling. Despite these limitations, the BiPathCPG proves a compelling alternative to CPG actuators for legged robots. Future work includes more durable designs with SLA resin suns [6], metal implementations, rigorous endurance testing, and analogous load-sharing in other two-stage topologies such as the Wolfrom gearbox.
References
- [1]P. López García et al., “Compact gearboxes for modern robotics: A review,” Frontiers in Robotics and AI, vol. 7, 2020.DOI
- [2]N. Kashiri et al., “An overview on principles for energy efficient robot locomotion,” Frontiers in Robotics and AI, vol. Volume 5 - 2018, 2018. [Online]. Available: https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2018.00129DOI
- [3]Y. Chi, Q. Liao et al., “Demonstrating berkeley humanoid lite: An open-source, accessible, and customizable 3d-printed humanoid robot,” 2025. [Online]. Available: https://arxiv.org/abs/2504.17249
- [4]A. M. Felix Grimminger et al., “An open torque-controlled modular robot architecture for legged locomotion research,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3650–3657, apr 2020. [Online]. Available: https://doi.org/10.1109/LRA.2020.2976639
- [5]W. Roozing and J. Volbeda, “Anti-backlash mechanisms for cycloidal drive robotic actuators: Design and evaluation,” IEEE Robotics and Automation Letters, vol. 10, no. 11, pp. 12 221–12 228, 2025.
- [6]K. Urs, C. E. Adu, E. J. Rouse, and T. Y. Moore, “Design and characterization of 3d printed, open-source actuators for legged locomotion,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 1957–1964.arxiv.org/abs/2202.12395
- [7]R. Bezzini et al., “Design and experimental evaluation of multiple 3d-printed reduction gearboxes for wearable exoskeletons,” Robotics, vol. 13, no. 11, p. 168, 2024.DOI
- [8]P. M. Wensing, A. Wang, S. Seok, D. Otten, J. Lang, and S. Kim, “Proprioceptive actuator design in the mit cheetah: Impact mitigation and high-bandwidth physical interaction for dynamic legged robots,” IEEE Transactions on Robotics, vol. 33, no. 3, pp. 509–522, 2017.DOI
- [9]B. Katz, J. D. Carlo, and S. Kim, “Mini cheetah: A platform for pushing the limits of dynamic quadruped control,” in 2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 6295–6301.DOI
- [10]T. Zhu, “Design of a highly dynamic humanoid robot,” PhD dissertation, University of California, Los Angeles, 2023, proQuest ID: Zhu_ucla_0031D_21632, Merritt ID: ark:/13030/m5tj684q. [Online]. Available: https://escholarship.org/uc/item/0qz3p57g
- [11]C. A. Pérez-Díaz et al., “Design and experimental characterisation of a novel quasi-direct drive actuator for highly dynamic robotic applications,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 183–189.
- [12]A. Singh, N. Kashiri, and N. Tsagarakis, “Design of a quasi-direct-drive actuator for dynamic motions,” Proceedings, vol. 64, no. 1, 2020. [Online]. Available: https://www.mdpi.com/2504-9364/1/1/1DOI
- [13]G. Bledt, M. J. Powell, B. Katz, J. Di Carlo, P. M. Wensing, and S. Kim, “Mit cheetah 3: Design and control of a robust, dynamic quadruped robot,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 2245–2252.
- [14]J. Choe, G. Kim et al., “Design of a 3-dof hopping robot with an optimized gearbox: An intermediate platform toward bipedal robots,” 2025. [Online]. Available: https://arxiv.org/abs/2505.12231
- [15]R. Bezzini, G. Bassani, C. A. Avizzano, and A. Filipeschi, “Design and examination of compound planetary gearboxes to investigate the impact of 3d-printed embedded bearings for assistive exoskeletons actuators,” Journal of Mechanical Design, vol. 148, no. 11, p. 113501, 2026.DOI
- [16]F. Yi, C. Zhang, S. Chen, S. Qiu, H. Zhang, Q. Du, W. Wang, and G. Yang, “A compact actuator for lower-limb exoskeletons with high torque density and high backdrivability,” IEEE/ASME Transactions on Mechatronics, vol. 31, no. 1, pp. 686–697, 2026.
- [17]A. Zhu, Y. Tanaka, F. Rafeedi, and D. Hong, “Cycloidal quasi-direct drive actuator designs with learning-based torque estimation for legged robotics,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 1–8.arxiv.org/abs/2410.16591
- [18]T. Yoshida, G. Endo, A. Okubo, and H. Nabae, “Experimental evaluation of a quasi-direct-drive actuator with a 3d-printed planetary gear reducer,” in 2023 IEEE/JS International Symposium on System Integration (SII), 2023, pp. 1–6.DOI
- [19]G. R. Wesley Roozing, “Experimental comparison of pinwheel and non-pinwheel designs of 3d-printed cycloidal gearing for robotics,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 7091–7098.DOI
- [20]W. Roozing and G. Roozing, “3d-printable low-reduction cycloidal gearing for robotics,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 1929–1935.DOI
- [21]R. Bezzini, G. Bassani, C. A. Avizzano, and A. Filipeschi, “Design of a novel flat cycloidal-planetary series reducer and experimental assessment for assistive exoskeletons actuation-cyclo,” IEEE Access, vol. 13, pp. 202619–202631, 2025.DOI
- [22]R. Bezzini, G. Bassani, C. A. Avizzano, and A. Filipeschi, “Nested planetary-cycloidal transmissions design for compact and lightweight exoskeletons actuators,” Reports in Mechanical Engineering, vol. 7, no. 1, pp. 253–272, 2026.DOI
- [23]A. Singh, D. Kapa, S. Joshi, and S. Kolathaya, “Compact: Computational optimization and automated modular design of planetary actuators,” 2025. [Online]. Available: https://arxiv.org/abs/2510.07197
- [24]M. Wikło, R. Król, K. Olejarczyk, and K. Kołodziejczyk, “Output torque ripple for a cycloidal gear train,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 2019.DOI
- [25]Y. Ding and H.-W. Park, “Design and experimental implementation of a quasi-direct-drive leg for optimized jumping,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 300–305.DOI
- [26]Q. Liao, B. Zhang, X. Huang, X. Huang, Z. Li, and K. Sreenath, “Berkeley humanoid: A research platform for learning-based control,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21781
- [27]H. Matsuki, K. Nagano, and Y. Fujimoto, “Bilateral drive gear—a highly backdrivable reduction gearbox for robotic actuators,” IEEE/ASME Transactions on Mechatronics, vol. 24, no. 6, pp. 2661–2673, 2019.DOI
- [28]G. M. Maitra, Handbook of Gear Design, 2nd ed. New Delhi, India: Tata McGraw-Hill, 1994.
- [29]H. Li, J. Liu, J. Ma, and Y. Shao, “Effect of the radial support stiffness of the ring gear on the vibrations for a planetary gear system,” Journal of Low Frequency Noise, Vibration and Active Control, vol. 39, no. 4, pp. 1024–1038, 2020. [Online]. Available: https://doi.org/10.1177/1461348419844642
- [30]Y. Xiong, K. Huang, F. Xu, Y. Yi, M. Sang, and H. Zhai, “Research on the influence of backlash on mesh stiffness and the nonlinear dynamics of spur gears,” Applied Sciences, vol. 9, no. 5, 2019. [Online]. Available: https://www.mdpi.com/2076-3417/9/5/1029DOI
- [31]R. C. Hibbeler, Mechanics of Materials, 11th ed. Pearson, 2022.
