Physical Motion Fidelity Metric

• Physical Motion Fidelity is a quantitative metric that defines realism by measuring the minimal L2 adjustment required to correct synthesized human motion.
• It uses reinforcement learning and physics simulators to evaluate kinematics, dynamics, and contact balance in motion sequences.
• PMF metrics provide a benchmark for training and assessing generative models, bridging the gap between simulation accuracy and perceptual realism.

Physical Motion Fidelity (PMF) Metric refers to a class of quantitative approaches designed to evaluate the realism and feasibility of generated human motion by measuring the degree to which synthesized trajectories conform to physical laws and produce plausible, executable movements within a physical simulation environment. Rather than relying solely on subjective human perception or heuristic rule-based checks, PMF metrics yield continuous-valued scores or fine-grained reward signals that reflect the physical validity of motion across kinematic, dynamic, and contact-balance dimensions. PMF serves as a foundational component for benchmarking human motion synthesis, training generative models with reinforcement learning, and bridging the gap between physically sound simulation and perceptual realism.

1. Motivation and Definition

Existing evaluation paradigms for human motion generation have predominantly relied on either coarse human perceptual judgments or rule-based physical plausibility checks. Such approaches present intrinsic limitations: human annotations tend to be subjective and granular only in a binary sense (“realistic” vs. “unrealistic”), while hand-crafted physics heuristics (e.g., foot-skating, joint limit violation) capture only isolated artifacts and do not aggregate into a global measure of realism.

Physical Motion Fidelity addresses these gaps by offering an objective, continuous, and simulator-grounded metric. For a motion sequence $x$, the physical fidelity is operationalized as the minimal L2 modification $\|x - x'\|_2$ required for $x$ to become physically valid, where $x'$ is the corrected trajectory found by a physical correction network $F_p$ trained via reinforcement learning to produce motions executable by a reference physics simulator (Zhao et al., 11 Aug 2025). Alternative paradigms, such as the PhyMotion reward, recover full 3D body meshes from video, retarget them to a physical simulation (e.g., MuJoCo), and quantify fidelity by aggregating physically meaningful violation terms along kinematics, balance, contact, and dynamics axes (Huang et al., 14 May 2026).

2. Methodological Frameworks for PMF Computation

Two principal frameworks have been established for defining and computing PMF metrics: the “minimal correction” approach and the “structured violation aggregation” approach.

Minimal Correction Approach

• For a given input $x$, PMF is quantified as the L2 norm between $x$ and its physically-corrected version $x'$:

$e_p = \|x - x'\|_2$

where $x'$ is computed by a learned correction function $\|x - x'\|_2$0, constrained by a physics simulator (e.g., IsaacGym), and trained with reinforcement learning rewards that penalize physical invalidity and excessive deviation from $\|x - x'\|_2$1 (Zhao et al., 11 Aug 2025).

Structured Violation Aggregation Approach

• For video-based or mesh input, the motion is retargeted to a controlled humanoid in a physics engine (e.g., MuJoCo) and scored along three axes:
  • Kinematic Feasibility ($\|x - x'\|_2$2): penalizing excessive joint speeds, self-penetrations