Torque-Balanced Procedural Locomotion for Physically Simulated Ragdoll Bipeds in Unity 3D: Architecture, Theoretical Analysis, and an Evaluation Protocol for a Real-Time, Non-Learned Control System

Physically simulated bipedal characters typically achieve balance and locomotion through one of twoestablished routes: offline optimization against a biomechanical or task objective, or a learned control policytrained with reinforcement learning against motion-capture reference data. Both routes require substantialoffline compute, reference data, and tooling external to the target runtime — resources not generallyavailable to a solo developer or small team building inside a commodity real-time game engine. This paperdocuments, theoretically analyzes, and defines an evaluation protocol for an alternative, minimalarchitecture: a hand-tuned proportional–derivative (PD) orientation controller for root balance, coupled toan independent, geometrically-triggered per-foot inverse-kinematics solver, both operating on a skeletonwhose physical plausibility is enforced entirely by engine-native ConfigurableJoint constraints rather than acustom dynamics model or offline solver.We situate this system within the broader history of physically-based character control, from earlybiomechanically-driven simulation through PD-based pose-graph controllers, offline muscle optimization,and modern reinforcement-learning approaches, and derive a linearized second-order model of the balancecontroller that yields concrete, falsifiable predictions — natural frequency, damping ratio, and expectedrecovery-time trends — directly from the controller's own gain parameters. We further describe acompanion interactive subsystem (procedural hand reach, world anchoring, and a physical object-pushinteraction handoff) built under the same architectural philosophy, and document specific implementationsubtleties surfaced during code-level review.A full evaluation protocol, with reference C# instrumentation, is defined for empirically testing thetheoretical predictions: a perturbation-recovery sweep, a gain-space stability sweep, gait-regularitymeasurement across terrain conditions, and per-frame computational cost profiling. Two of the four plannedexperiments have been executed at the time of writing: a 48-trial perturbation sweep found peak tilt scalingapproximately linearly with impulse magnitude (R²=0.83) with zero falls and sub-degree tilt across thetested 5–30 N·s range — confirming the controller remained in the small-angle regime the linearizationassumes, but also revealing that this range was too conservative to characterize actual recovery dynamics orfailure boundaries; and a gait-logging dry run produced insufficient data (5 events, one terrain condition) tocompute the intended regularity statistic. The gain-space stability sweep and computational-cost profilingremain future wor