RLxUSD v0.1 A Minimal Convention for Representing 3D Reinforcement Learning Episodes in Universal Scene Description (USD)

We present RLxUSD v0.1, a minimal and extensible convention to encode complete reinforce-
ment learning (RL) episodes as OpenUSD scenes. The convention prescribes a small set of scene
primitives, a metrics: namespace for time-aligned time series, and an episode summary stored
in stage metadata with both required and optional fields. By leveraging USD timeSamples and
native extensibility, RLxUSD unifies geometry, metrics, and metadata in a single, inspectable
artifact. We provide a compact reference implementation, integrations with Gymnasium and
Stable-Baselines3 (SB3), a usdview heads-up display (HUD) for synchronized visualization, and
a small reproducible dataset across Random, Greedy, and PPO agents in 16×16 and 32×32
gridworlds. The result improves transparency, portability, and reproducibility for episode-level
RL research and engineering.