Creating manufacturable blueprints for coarse-grained virtual robots

Zihan Guo; Muhan Li; Shuzhe Zhang; Sam Kriegman

cs.ROMar 13, 2026

Creating manufacturable blueprints for coarse-grained virtual robots

Authors: Zihan Guo, Muhan Li, Shuzhe Zhang, Sam Kriegman

Organizations: Northwestern University

Abstract

Over the past three decades, countless embodied yet virtual agents have freely evolved inside computer simulations, but vanishingly few were realized as physical robots. This is because evolution was conducted at a level of abstraction that was convenient for freeform body generation (creation, mutation, recombination) but swept away almost all of the physical details of functional body parts. The resulting designs were crude and underdetermined, requiring considerable effort and expertise to convert into a manufacturable format. Here, we automate this mapping from simplified design spaces that are readily evolvable to complete blueprints that can be directly followed by a builder. The pipeline incrementally resolves manufacturing constraints by embedding the structural and functional semantics of motors, electronics, batteries, and wiring into the abstract virtual design. In lieu of evolution, a user-defined or AI-generated ``sketch'' of a body plan can also be fed as input to the pipeline, providing a versatile framework for accelerating the design of novel robots.

Figures & tables

Fig. 1: Converting high-level abstractions into real robots. The pipeline begins in simulation (virtual stage; A-C ) where abstract, voxelized robots—“virtual creatures”—can freely evolve complex endoskeletal morphology without recourse to real world manufacturing constraints. A design with high fitness in simulation is shown in A. This voxelized representation is fed as input to the “semi-virtual” stage for processing ( D-F ). Here, the rigid bone processor converts rigid voxels to bone meshes (gray surfaces in D) and cuts cylinders (green rolls in D) along the joint axis to reserve space for motors. The soft skin processor cleans space around joints (green spheres in E), applies erode-dilate smoothing and converts soft voxels into hollow skin meshes (blue surfaces in E) to avoid self-collision. The output of these processors is composed into a layered data structure (F) and forward propagated to the next stage. We refer to the next part of the pipeline the “semi-real” stage because it uses solver nodes to incrementally resolve manufacturability constraints, but the interim design is neither realized nor a fully fledged blueprint. First, admissible joint motion ranges are estimated by the motor solver ( G ) and candidate motor holder/connector depths are scanned along each joint axis to select the optimal motor position ( H ). Next, the electronics solver carves cavities for the controller and battery installation box into the two body parts with the largest volume ( I ). Wire tunnels are then routed ( J ) through both rigid and soft meshes by the wire solver. The resulting assembly-ready multi-material body parts are then 3D printed and assembled in the real world ( K ). A cross-section of one of the endoskeletal body parts is displayed (in K). The evolved tripedal design successfully transferred to reality, exhibiting forward locomotion under the learned closed-loop policy ( L-O ). The body rocks back and forth, pivoting on single hind limb as the left and right forelimbs alternate contact with the ground. The trajectory of the right forelimb is traced (from purple to yellow) in O.

Fig. 2: Running the pipeline. We tested our pipeline against a large, diverse set of designs randomly sampled from the latent space, a representative subset of which are displayed ( A-R ). For each design, motors (green cylinders), a battery (blue boxes), and a controller box (red boxes) are automatically positioned and connected by wire tunnels (yellow lines). The resulting anatomy is then assigned a manufacturability score based on the estimated feasibility/strength of the selected motor positions, the availability of free space for the electronics and battery, the curvature necessary for cable tunnels, the clearance above the electronics box, and collision of body parts after motor integration. Low-quality designs have lower scores. For example, designs that have very long cable tunnels (I), are too thin for the electronics box (M, N), lack the spacing between joints for the motors (K, L), or possess an invalid kinematic tree (Q, R).

Fig. 3: Failure modes. Designs were sampled randomly from the latent space, which can be projected by linear discriminant analysis into a 2D map ( A ). Each axis is a linear combination of latent features, and shows that failures from the motor ( B ) and electronics ( C ) solvers and the rigid bone processor ( D ) can be isolated and avoided during design evolution. A histogram of the manufacturability scores from sampled designs reports the number of robots within each score interval ( E ).

Fig. 4: The electronics system. The selected mechatronics components are shown. The system was designed to be modular and extendable. For example, if virtual creatures with hearing or vision are to be realized in future work, microphones and cameras could be easily incorporated into this system with off-the-shelf modules.

Solver stage	Remaining ratio
Motor solver	89.24%
Electronics solver	77.85%
Wire solver	99.87%
Complete pass-through success	66.96%

TABLE I: Remaining ratio after each stage for randomly sampled morphologies (N=1000).

Fig. 5: Sketch2robot. Abstract design representations are also convenient for human designers, and the pipeline readily accepts a hand-drawn “sketch” in voxel space as input ( A ). Voxel workspaces in particular can provide an intuitive and enjoyable interface, as evidenced by the popularity of games like Minecraft. Sketches input to the pipeline are automatically processed ( B ) and embedded with hardware to produce a CAD file ( C ) that can be directly printed and assembled ( D ).

Robotics

Robotics studies the design and control of machines that perceive, move, and act in physical environments.

See all 751 papers

Draft: A Parametric Tool for Robot Design Exploration
Sep 29, 2026David Nguyen, Marcelo Coelho, Sangbae KimRobot Morphology-Control Co-Design Robotics Simulation
Department of Mechanical Engineering, Massachusetts Institute of Technology · Department of Architecture, Massachusetts Institute of Technology
Agentic Robot Creation through Generation, Verification, and Revision
Oct 13, 2026Xuebing Li, Yibo He, Sitong Liu +2Agentic AI for Robotics Robotics
ShanghaiTech University
Bridging Language and Physics: Automated Design of Continuum Robots with Large Language Models
Sep 8, 2026Jingyi Chen, Mohan Zhang, Laura Yao +5Robotics Continuum Robot Design
School of Computer Science and Technology, University of Science and Technology of China, Hefei, China · Department of Computer Science, University of North Carolina at Chapel Hill, Chapel Hill, USA · Department of Computer Science, University of Central Florida, Orlando, USA

Sep 29, 2026J/K move · Enter open · S save

Draft: A Parametric Tool for Robot Design Exploration

David Nguyen, Marcelo Coelho, Sangbae Kim

Department of Mechanical Engineering, Massachusetts Institute of Technology · Department of Architecture, Massachusetts Institute of Technology

Robot Morphology-Control Co-Design Robotics Simulation Robotics

Open paper AskPDF arXiv

Robot performance is often limited by the cost of iterating on morphology and control together, since every computer-aided design (CAD) change has to be carried into a simulation-ready model before control work begins. Co-design methods attempt to close this gap, but each uses a model generator written for a single platform or lack the use of real-world data to suggest that designs are plausible. We present Draft, a parametric generation tool whose generalized engine compiles any parametric tree of serial chains into a simulation-ready MJCF model, without CAD. It allows engineers to explore design tradeoffs through easily adjustable models and evaluate how changes influence controller performance. Draft grounds the free parameters of each design using trends fitted to a survey of

114

actuators and

49

published robot descriptions, so that a generated robot is anchored to real-world hardware. We validate those trends wholistically by building twins of four off-the-shelf robots, whose masses agree to

1.10\times

geometric mean fold error. Finally, we demonstrate how Draft exposes design tradeoffs by evaluating three quadrupeds through a two-stage reinforcement learning curriculum.

Abstract

Robotics

Explore similar work

Draft: A Parametric Tool for Robot Design Exploration

Agentic Robot Creation through Generation, Verification, and Revision

Bridging Language and Physics: Automated Design of Continuum Robots with Large Language Models