[Paper Notes] Draft: A Parametric Tool for Robot Design Exploration
Published:
This post supports English / 中文 switching via the site language toggle in the top navigation.
TL;DR
Changing a robot’s leg length also changes its geometry, mass distribution, inertia, and actuator requirements. Draft makes those dependencies explicit: a parameter file, an expression-based kinematic tree, and a root-body mesh specification compile into a simulation-ready MuJoCo MJCF model. Fits to real actuators and robot descriptions supply plausible physical parameters. Engineers can then train controllers and compare designs before investing in detailed CAD.
“Automatic robot skeleton generation” is a useful starting description. More precisely, Draft generates a physical model from a supplied body specification and evaluates its capabilities. The designer still chooses the topology and the next design to try; an outer design optimizer is future work.
Paper and sources
Draft: A Parametric Tool for Robot Design Exploration, by David Nguyen, Marcelo Coelho, and Sangbae Kim, MIT. These notes follow the September 29, 2026 preprint, arXiv:2609.38405v1, and its eight-page PDF. The authors provide official code and datasets and a supplementary video. Results below are author-reported; the practical interpretation at the end is my reading of the work.
1. How a skeleton becomes a simulation model
The parameter file supplies quantities such as link lengths, densities, and motor classes. Fields in the kinematic tree contain arithmetic expressions referencing those parameters. A link radius can therefore depend on its motor’s dimensions, and a torso mesh can change with shoulder width. Editing one parameter propagates through the dependent visual, collision, inertial, and actuation properties.
The generator assembles four primitives. A basic link is a structural cylinder with the motor for the child joint at its distal end. A sphere link ends a chain with a contact point; a foot link supplies a support surface. The root link is a closed mesh lofted from cross-section polygons and planes. These primitives share a compiler across humanoid and quadruped specifications. Fixed joints also allow combined modules to place a knee actuator near the hip. Parallel mechanisms are outside the current link library.
MuJoCo’s inertiafromgeom computes primitive-link inertias from geometry and density; trimesh supplies root-mesh inertial properties. MJCF directly represents gear ratios and joint armature, which makes the actuator model part of the generated dynamics. The paper reports compilation to validated MJCF in 50 ms for a 12-joint quadruped and 202 ms for a 29-joint humanoid on a laptop CPU. These are model-generation times; controller training is a separate cost.
2. Torque and speed determine the rest of the actuator
A simulator will accept physically implausible combinations of torque, mass, and size unless the generator constrains them. Draft surveys 114 integrated actuator modules, including motor, transmission, bearings, and housing. Its catalog contains 33 quasi-direct-drive, 33 MidGear, and 48 high-reduction modules. The fitted relationships make actuator choices carry their physical costs.
The inputs are peak output torque $\tau$, no-load output speed $\omega_{\mathrm{NL}}$, and aspect ratio $a=\ell/(2r)$. Table I gives the following relationships, using SI quantities: torque in N·m, speed in rad/s, mass in kg, volume in m³, dimensions in m, and inertia in kg·m².
\[\begin{aligned} N &= 327\,\omega_{\mathrm{NL}}^{-1.066},\\ m &= 0.0653\,\tau^{0.666},\\ V &= \frac{\tau^{0.700}N^{-0.167}}{23948},\\ r &= \left(\frac{V}{2\pi a}\right)^{1/3},\qquad \ell=2ar,\\ J &= 27.2\,r^4,\qquad J_{\mathrm{armature}}=N^2J. \end{aligned}\]Here $N$ is reduction ratio and $J$ is rotor inertia. These are empirical fits with unit-dependent coefficients. Most rows use 103 geared modules; the rotor-inertia fit uses the 88 entries that publish that quantity. The full 114-entry survey is larger than either fitting subset.
The last expression matters for control. A high gear ratio increases the rotor inertia reflected at the joint quadratically. Choosing a slow, highly geared actuator therefore changes joint dynamics as well as the available torque and speed. During learning, a DC-motor model enforces a linear torque–speed curve. PD gains follow the hand-tuned scaling $k_p=1.43\tau_{\mathrm{peak}}$ and $k_d=0.071\tau_{\mathrm{peak}}$.
Leave-one-out geometric mean fold errors are 1.28× for mass, 1.33× for reduction, 1.30× for volume, and 1.56× for rotor inertia. The inertia estimate is the loosest: housing radius is only a proxy for rotor radius. These relations provide a baseline for integrated commercial modules; custom actuator designs can fall outside it.
3. Structural mass is fitted separately
The authors inspect 118 published robot descriptions and retain 29 humanoids and 20 quadrupeds, totaling 1,037 segments. To estimate structural mass, they subtract an estimated actuator mass from each link, then fit
\[m_{\mathrm{structure}}=cL^e\]by segment class and body type, with $e\geq0$. This separation avoids counting the motor twice. It also exposes an assumption: the correct link must be assigned each actuator’s mass. Negative residual masses are excluded. Material choices and errors in actuator-mass estimation contribute additional variation.
Humanoid and quadruped limbs receive separate fits. For example, Table II gives thigh exponents of 1.36 and 2.69 respectively. Torsos use volumetric density: 716 kg/m³ for humanoids and 654 kg/m³ for quadrupeds. Held-out whole-robot mass errors are 1.22× and 1.18×; Table II’s individual segment errors are in-sample errors and should be read separately.
A design filter rejects actuators outside catalog/transmission-family extremes and warns about unusual aspect ratios, structural densities, and actuator mass fractions. Passing this filter establishes consistency with the surveyed range. Detailed mechanical engineering remains necessary to turn the model into hardware.
4. What the validation establishes
Draft rebuilds Unitree G1, H2, Go2, and B2 using vendor geometry and joint capabilities while deriving mass, density, inertia, and armature. Generated total masses are 111%, 91%, 101%, and 84% of the reference masses, giving a geometric mean fold error of 1.10× across these four robots.
The dynamics comparison is less close. Across 300 configurations, Go2 has median joint-space mass-matrix error of 12.8% and gravity-vector error of 21.4%. Humanoid errors are larger, partly because some joint axes are misaligned; holding those axes fixed reduces the discrepancy. Rotor armature is disabled in this comparison because only G1 publishes per-joint armature. Thus the experiment does not validate every generated actuator-inertia term.
The locomotion study generates three quadrupeds and trains each with PPO in mjlab, using ten independent seeds. Stage one learns velocity tracking on mildly uneven ground with occasional pushes. Stage two branches into three policies, progressively increasing commanded speed, terrain height, or push magnitude. Rewards scale with leg length using Froude similarity: for $s=L/(0.48\,\mathrm m)$, lengths scale with $s$, while characteristic speed and time scale with $\sqrt{s}$.
| Design | Mass | Leg length | Peak speed | Step height | Recoverable velocity kick |
|---|---|---|---|---|---|
| Cheetah, QDD 6:1 | 33 kg | 0.48 m | 3.63 m/s | 0.12 m | 7.9 m/s |
| Bear, MidGear 20:1 | 110 kg | 0.74 m | 4.33 m/s | 0.12 m | 8.6 m/s |
| Giraffe, high reduction 80:1 | 91 kg | 1.30 m | 2.17 m/s | 0.20 m | 5.3 m/s |
These are Table V medians over ten seeds, each based on 256 rollouts. “Kick” is an imposed velocity change, not a force. Bear leads absolute speed and kick recovery; Giraffe clears the tallest step. After normalizing speed and kicks by $\sqrt{gL}$ and step height by $L$, Cheetah leads those relative measures. The choice of metric changes which design looks preferable.
The paper also reports drivetrain efficiency from mechanical output and modeled copper loss. This metric excludes a full accounting of battery and auxiliary power. Post-training link-mass perturbations modeled on Go2/B2 reconstruction residuals preserve the reported capability trends without retraining.
5. Where I would use Draft
I would use Draft to narrow early choices about proportions, actuator classes, and motor placement before detailed CAD. Its most useful feature is the coupling between decisions: increasing a torque requirement brings corresponding mass, geometry, and inertia changes into the controller experiment. That makes comparisons more informative than changing a torque limit alone.
The terrain experiment also shows why a learned policy needs inspection. Some policies stop attempting harder steps while actuator utilization remains low, because surviving earns more reward than trying and falling. A measured terrain limit can therefore reflect the learning setup. Even actuator saturation under a learned gait does not establish a globally optimal controller.
For a new morphology, I would first check whether the actuator and structural fits cover it, then examine joint utilization and sensitivity to mass estimates before trusting the ranking. The current evidence is strongest for early exploration of the surveyed legged robots. Extending the method to dexterous hands or parallel mechanisms requires additional data or model support.