A robot that has to cross a busy Zürich Hauptbahnhof concourse without frightening anyone cannot be trained on the concourse — you get one shot, and the failure mode is a person on the floor. So the training moves indoors, into a simulator. The frontier question is whether the simulated crowd is faithful enough that a policy learned against it survives contact with real pedestrians. A new arXiv paper, SONG: A Photorealistic 3D Gaussian Simulation Platform for Benchmarking Social Navigation (arXiv:2607.25219v1, cs.RO), answers with a platform built entirely on 3D Gaussian splatting (3DGS) — the radiance-field representation introduced by Kerbl et al. at SIGGRAPH 2023 out of Inria’s GRAPHDECO group, which renders a scanned scene as millions of oriented Gaussian blobs instead of triangles.

The mechanism is worth reading closely, because it is the same stack an architecture office will use to rehearse a fabrication cell or a public interior. SONG represents both the scene and the human avatars as Gaussians, then drives each pedestrian along a trajectory a large language model proposes from the scene’s semantics, and fills in the full-body walking motion with a trajectory-conditioned generator for continuous, natural movement. On top of it sits SONG-Bench, a set of evaluation episodes stratified by difficulty, scored against a multi-dimensional metric suite that measures three axes separately: effectiveness, safety, and social compliance.

Three findings land hard. Vision-based social navigation is far from solved. A safety deficit precedes any etiquette deficit — the robots fail at not-colliding well before they fail at being polite. And, most usefully for anyone buying compute, real-world data mattered more than model scale: fine-tuning on SONG’s curated real episodes lifted real-world success rate where a bigger model alone did not. That last point rhymes with the GPU-parallelised ManiSkill3 line of work (arXiv:2410.00425, from Hao Su’s lab at UC San Diego) — throughput buys you samples, but the samples have to be honest.

←TODAY: A robotics lab renders a scanned building as Gaussians and trains a navigation policy inside it — the digital twin becomes a rehearsal stage. →3012: Every public interior ships with a certified splat-twin, and the machines that will move through it are pre-cleared against it before they arrive. Fulcrum: The twin is only trustworthy if the scan of a public building is held under the same residency rules as the building’s other records.

Here is the Swiss seam. A splat-twin of a Kantonsspital lobby or a Bahnhof is a high-resolution recording of a public space and the people in it. The moment a municipality procures a simulation platform to certify a delivery robot or a cleaning machine, it is procuring a data pipeline — and the SONG design, which the paper shows needs real episodes more than raw scale, means that pipeline wants your local footage. That is a data-residency question dressed as a robotics tool.

Atelier: An office trialling any Unity/Unreal or Gaussian-splat digital-twin workflow for fabrication or facility rehearsal is quietly becoming a data controller. Monday move: before you scan a live public interior into a twin, write one clause into the vendor or internal spec — the raw scan and every derived splat stay on infrastructure you name, in a jurisdiction you name — and treat any tool that cannot honour it as failing the tender, not as a feature gap.

Hack: Score SONG’s safety-before-etiquette finding yourself with a one-line clearance metric — the minimum distance between your robot’s path and each pedestrian trajectory, sampled step-by-step. If that minimum drops below your body-buffer, the run is unsafe no matter how polite it looked. This is the safety axis of SONG-Bench reduced to a single number your desk can compute before trusting any twin.

import numpy as np
robot = np.load("robot_xy.npy")      # (T,2) path samples, metres
ped   = np.load("ped_xy.npy")        # (T,2) one avatar's trajectory
clearance = np.linalg.norm(robot - ped, axis=1).min()
print(f"min clearance {clearance:.2f} m — unsafe" if clearance < 0.5 else "ok")

Half a metre is a defensible personal buffer; tighten it for a corridor, loosen it for a plaza. The number, not the vibe, is the gate — which is exactly SONG’s point about safety preceding etiquette.

Read the SONG paper for the metric definitions, then read the data-residency clause in your next simulation-platform tender. If the clause is missing, write it in before you scan a single public room.

Source: arXiv