Using scenes in simulators | Docs

Every segment carries the scene it was recorded in: Gaussian splats for appearance and meshes for collision, in the site's world frame at metric scale. Export them to evaluate policies in the places your robots actually work.

Isaac Sim and Isaac Lab

Python

scene = client.segments.get("seg_4f0a").scene_at(t=0.0)
scene.to_usd("kitchen_lab2.usdz")

Or export a whole slice:

Shell

rayform export slc_7c1e04aa --format usd --out ./scenes/

The USD file contains:

Prim	Contents
`/World/Background`	Gaussian splats of the static scene
`/World/Collision`	Collision meshes, with convex decomposition for physics
`/World/Objects/*`	Moved objects as separate prims, posed at time `t`
`/World/Cameras/*`	The recording's cameras with their intrinsics

MuJoCo

Python

scene.to_mjcf("kitchen_lab2/")   # writes scene.xml plus meshes/

Python

import mujoco

model = mujoco.MjModel.from_xml_path("kitchen_lab2/scene.xml")
data = mujoco.MjData(model)

MJCF holds the collision geometry only; MuJoCo does not render Gaussian splats.

Posing objects at any moment

Objects that moved during the recording are separate entities with trajectories. scene_at(t) places them where they were at time t, so you can start a simulation from any moment of a real episode:

Python

seg = client.segments.get("seg_4f0a")
before = seg.scene_at(t=seg.events()[0].t_ns - 500_000_000)   # half a second before the first grasp