RoboCompiler: Graph-Native Compilation of Closed-Chain Robots for Consistent Modeling, Control, and Simulation

Closed-chain machines are common in industry and awkward in software. This paper compiles one mechanical interface that both the controller and the simulator consume, so the two can no longer disagree.

Background

Robots with kinematic loops, coupled actuators and changing contacts demand consistent models of configuration, motion, force and dynamics. In practice those interfaces are often reconstructed separately for the control stack and for the simulator, which makes closure and actuation consistency difficult to maintain and easy to break. The result is a quiet class of bugs: control and simulation no longer agree about the same mechanism, and every revalidation has to chase the mismatch through two divergent code paths.

Core Innovation

RoboCompiler is a graph-native framework that compiles a canonical mechanism graph into a shared mechanical interface. From bodies, joints, frames, inertias and actuator ports it constructs closure paths and analytic residual Jacobians, then assembles feasible configurations through rank-checked continuation and correction. A tangent lift maps independent velocities to full robot and task motion, while paired actuator-port maps preserve virtual work. A constraint-curvature correction extends the reduction to accelerations and projected rigid-body dynamics, including floating-base and support modes. Cycle-local evaluation, generated Jacobians and dependency-aware reuse allow localized updates when closure inputs change, so a single compiled interface is reused across control and simulation instead of being rebuilt for each.

Results

• Evaluated on a Komatsu excavator, a Unitree Go2, a Franka Panda, Kangaroo and a six-UPS Stewart platform.
• High-precision constrained-dynamics comparisons and independent Pinocchio checks confirm mechanical consistency.
• MuJoCo and Isaac Sim/PhysX executions demonstrate task performance and model reuse under native contact.
• On Kangaroo, compilation reduces residual-and-Jacobian evaluation time by 96.7% and closed-loop rollout wall time by 66.8%, with dynamics and control held fixed.

Limitations

The efficiency headline is measured on Kangaroo and holds dynamics and control fixed, so it isolates the modeling layer rather than end-task performance; the abstract does not report comparable speedups for the other platforms. The consistency claims rest on checks against Pinocchio and on simulator execution rather than on field deployment, and no hardware results are reported. Whether dependency-aware reuse still pays off when many closure inputs change at once, or under highly dynamic contacts, is not addressed in the abstract, nor is the cost of the initial compilation for very large mechanisms.

Industry Implications

Closed-chain and loop-closed machines - excavators, parallel platforms, legged robots with coupled transmissions - are widespread in industry and genuinely awkward in software: teams maintain two divergent models of the same mechanism and pay for it in continuous revalidation. A single compiled mechanical interface that both the controller and the simulator consume is a maintainability win before it is a performance one, because consistency bugs between the two are expensive to find and easy to hide. The reported 96.7% and 66.8% reductions on Kangaroo point to a concrete payoff in simulation-heavy workflows such as reinforcement-learning training and model-predictive control, where the same Jacobians are evaluated millions of times and the loop dominates wall-clock cost. For platforms that must document their behavior, a shared interface backed by mechanical-consistency checks is also an easier artifact to trust than a hand-rebuilt pair.