\section{Method} \label{sec:method} \sys has an offline part (Python, PyTorch, gsplat) that turns nuScenes logs into 3DGS scenes and auxiliary maps, and an online part (C++, Vulkan) that renders six cameras, simulates the ego vehicle and exposes observations to the policy. \subsection{Reconstruction} \label{sec:recon} Our target set is the 284 daytime Singapore scenes of nuScenes (383 Singapore scenes minus 99 night scenes; all Singapore rain scenes are at night), reconstructed from all six cameras at 12\,Hz including sweeps, with every 8th frame held out. \TODO{the batch has not been run yet; numbers below are from pilot scenes.} Gaussians are initialized from accumulated lidar sweeps (dynamic boxes removed, 0.15\,m voxels, colored by projecting into the keyframes) and optimized with MCMC densification~\cite{kheradmand20243dgsmcmc}, a per-camera affine color correction, a learned refinement of the camera extrinsics, and an L1 loss between rendered expected depth and lidar depth maps re-projected with the refined extrinsics (weight $0.1$). Lidar supervision reduces the depth error against lidar from 22.0\% to 2.3\% at a cost of 0.8\,dB PSNR on the recorded trajectory, and visibly preserves curbs and lane markings after lane changes. Annotated agents are masked out of the photometric loss, giving an \emph{environment-only} model into which agents are inserted at simulation time. The sky is either modeled by a learned cube map with a SegFormer~\cite{xie2021segformer} sky-mask loss, as in OmniRe~\cite{chen2025omnire}, or, for models trained without it, we \emph{label} sky Gaussians after training: a Gaussian is sky if, among the training views in which it projects inside the image and is not behind the lidar surface ($z < 1.1\,d_\text{lidar}+0.5$\,m), more than 40\% land on sky pixels. Labeled Gaussians---typically bright floaters a few meters to tens of meters away that become visible from novel viewpoints---can then be dropped at render time. \subsection{Reconstruction at speed} \label{sec:speed} Pilot reconstructions revealed a strong dependence of quality on ego speed (\cref{tab:speed}). On the fastest scene, changing the densification strategy (MCMC vs.\ default), the depth loss or the extrinsic refinement leaves the full-length result low, while restricting training to a time window of 2.5--5\,s (30--60\,m of road) recovers most of the quality with either strategy, and the training loss falls by $2$--$3\times$: the full-length scene is under-fit rather than over-fit. 77\% of the daytime scenes (218 of 284) average more than 3\,m/s. We therefore split scenes into temporal chunks; stitching chunks and overlapping scenes into a map is discussed in \cref{sec:limitations}. \subsection{Real-time renderer} \label{sec:renderer} \paragraph{Render graph.} A frame is a JSON document listing pipelines, resources and passes. Each pass declares a list of $(\text{resource}, \text{usage})$ pairs, \eg \texttt{storage\_read}, \texttt{color\_attachment}, \texttt{transfer\_src}. Execution order is the declaration order, which also defines the access order, so dependency edges always point forward and no cycle can occur. A single traversal loop tracks, for every resource, the stage and access of its last write, the stages that have read it since and its image layout, and emits exactly the memory and layout barriers implied by the next usage. Pass kinds (compute dispatch, radix sort, buffer/image copy, raster with a list of draw items) are small executors; adding a pass, an image or a render target only edits JSON. Shader interfaces (push-constant size, descriptor types, compute local size, fragment outputs) are read from SPIR-V by reflection, and all pipelines share one bindless descriptor set. Post-processing effects (\eg, a camera ISP model with exposure, white balance, vignetting and sensor noise for robustness tests) can be authored as node graphs that are compiled to GLSL at build time. The current graph (\cref{fig:graph}) has about ten passes; the Khronos synchronization validator reports no hazards. \paragraph{Sorting and blending.} Splats are composited back-to-front, which requires a per-frame depth sort. Our baseline bitonic sort over all $2^{21}$ Gaussians needed 231 dispatches. Instead, the preprocessing pass compacts visible Gaussians with an atomic counter, and a four-pass LSD radix sort (8-bit digits, histogram $\rightarrow$ scan $\rightarrow$ stable scatter with subgroup-ballot ranking) sorts only the visible keys, reading the count from the indirect-draw buffer so that no CPU round trip is needed. Colors are blended into RGBA16F targets (PSNR 64\,dB against FP32), and a screen-space FXAA pass (0.14\,ms) smooths mesh and map-overlay edges; MSAA is avoided because it multiplies splat blending cost by the sample count. \subsection{Lane-shift distillation} \label{sec:distill} Following Difix3D+~\cite{wu2025difix3d}, we render training cameras translated laterally, correct the renders with the Difix single-step fixer conditioned on the real image of the same timestamp, mask annotated agents in the pseudo views, and fine-tune the scene with pseudo views sampled with probability 0.3 and weight 0.5. To keep pseudo targets consistent we proceed progressively ($\pm0.5\to\pm1\to\pm2\to\pm3$\,m), each stage rendering from the previous stage's model. \TODO{Fine-tuning currently uses the default densification strategy, which prunes the MCMC model from 2.0M to 0.9M Gaussians; re-run with MCMC fine-tuning and report extrapolation metrics.} \subsection{Closed loop} \label{sec:loop} The ego vehicle follows a kinematic bicycle model with longitudinal limits and suspension animation, rendered as a glTF vehicle with physically-based shading from image-based lighting baked from the scene. Collisions are tested against an obstacle grid built from static lidar points (0.25--2.5\,m above ground, road interior cleared) and against the nuScenes drivable-area boundary (curbs). Observations are the six camera images; the viewer also provides a chase camera, a bird's-eye view and a Tracy~\cite{tracy} profiling interface with per-pass GPU timestamps. \TODO{policy interface and HD-Score~\cite{zhou2024hugsim} evaluation.} \begin{figure}[t] \centering \includegraphics[width=\linewidth]{img/render_graph.pdf} \caption{\textbf{The frame as a render graph} (generated from the JSON). Boxes are passes in execution order, cylinders/sheets are buffers/images; edges are labeled with the declared usage from which barriers are derived.} \label{fig:graph} \end{figure}