LiteReality-Agent

An Agentic System for Code-Based Indoor Reconstruction from RGB-D Scans

TL;DR

We introduce the first open-source, end-to-end toolkit for reconstructing interactable indoor 3D scenes. The LiteReality Scanner app captures real-world rooms, while LiteReality-Agent transforms the scans into complete, interactive 3D scenes with realistic, graphics-ready appearances and articulated assets.

02

Look into a reconstruction

Interactive viewer for the reconstruction results.

Elliott Office · QC reconstructionopen full viewer ↗
drag orbitscroll zoomGroups hide / isolateSplit view real vs render
The scanner app

LiteReality

Our iOS LiDAR room scanner. Walk a room with an iPhone or iPad and it exports a single bundle — posed RGB frames, LiDAR depth, ARKit camera poses and the RoomPlan layout — everything the reconstruction pipeline needs.

iOS · iPadOS (LiDAR) — on the App Store
04

How it works

LiteReality-Agent starts by initializing the room from the scan through a deterministic pipeline, then an authoring stage refines it toward a realistic appearance. At its core the agent works by editing the scene's Room.py program — its entire workspace — in a render–compare–critique loop, with quality control gating the result.

input

Scan

Posed RGB + LiDAR from the app.

deterministic

Initialize

Walls, openings & cameras rebuilt into a blank room.

agentic loop

Author

Agent edits Room.py, refining against the capture.

gate

QC

Checks the scene meets quality standards.

output

Final scene

The realistic, interactive, articulated 3D result.

The authoring loop
edit Room.pyrendercomparecritique
Agent tools
render_and_compareselect_viewgridfetch_materialscriticcompile bashglobreadedit
Room.py

Programmable rooms. The whole scene — geometry, materials, articulation — is one Python program the agent edits.

Toolset

Purpose-built tools. A set the agent composes — render & compare, select views, measure, fetch materials, critique, compile — alongside the usual agent tools (bash, glob, read, edit) for everything else.

QC

Quality control. Automatic checks keep every output scene within standard before it ships.

05

A programmable scene · Room.py

The whole scene — geometry, materials, articulation — is defined in one Room.py program. That is what makes it the agent’s editable workspace: legible, replayable and portable.

Programmable

The agent edits lines of code, never raw meshes.

Deterministic

Rerun the file to rebuild the scene — assets are regenerated, not stored.

Bounded

A fixed set of methods, so every change stays legible and traceable.

Portable

The same file exports to Blender, glTF or a simulator.

06

Applications

Because the whole scene is an editable Room.py program, a plain-English prompt becomes a code edit — the room is programmable. Three things it can already do, each a clip from the same authored office:

APP 01

Text-based scene editing

Everything lives in a Room.py file, so the room is a programmable object — a typed request becomes a code edit. Re-texture the walls or swap the floor in a few lines: exposed brick, walnut panelling, sage-green paint. The geometry never moves.

Demonstrated
APP 02

Scene rearrangement

Move the furniture — nothing is deleted. From a prompt the agent re-lays-out the room while keeping everything grounded and collision-free: clear the floor for a standup, or seat eight around the table.

Demonstrated
APP 03

Decoration

Fixtures, props and banners, placed in 3D. The agent fetches real assets and writes procedural geometry into Room.py — pendant lamps, a birthday setup, a World‑Cup party — all clear of the door swing.

Demonstrated
APP 04

Intrinsic outputs

Because the scene is fully authored, it renders perfect intrinsics for free — pixel-exact semantic segmentation, depth, normals and albedo, from the very same cameras. Dense ground truth no real capture can hand you.

Demonstrated
On the roadmap
Plannedapp 05

Robotics & embodied AI

Export the articulated, collision-ready room to a physics simulator as a stage for training and evaluating embodied agents in real layouts.

07

Gallery of reconstructions

Every reconstruction, interactive in the browser — real capture beside the render, from the same camera. Click any scene to orbit it.

▸ Overview & walkthrough videos will be dropped in here. Each card links to the live, orbitable reconstruction with a real-vs-render comparison and per-group visibility toggles.

09

Citation

bibtexto be updated on release
@article{huang2026litereality,
  title   = {LiteReality-Agent: An Agentic Harness for
             Simulation-Ready Indoor Reconstruction},
  author  = {Huang, Zhening and Li, Yueyan and Chiu, Johnathan and
             Lyu, Xiaoyang and Zhou, Matt and Yao, Yuxin and
             Lasenby, Joan and Wu, Shangzhe},
  year    = {2026},
  note    = {Project page: litereality-agent}
}