KevinXu02 / aha-3d

# AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra
**Project page**
AHa-3D turns an ordinary indoor video into an editable 3D scene in Blender. It
rebuilds the room from reference geometry and a library of reusable furniture and
materials, estimates the motion of the people in the video, places them in the
shared room with foot-ground and object contact refinement, and renders the result
with the original camera and timing. It can also generate new actions for people
in the scene and export saved scenes as interactive browser demos.
## Demo
An office rebuilt from a 10-second video and rendered through the recovered source
camera (MP4).
Open the editable scene in examples/office48 with Blender 5.2. The project page
shows it next to the source video and hosts an interactive scene.
## Quick start
**Requirements:** one Linux workstation (Ubuntu or WSL2) with an NVIDIA GPU,
Python 3.11 and compatible Blender for scene assembly. Human skinning also
needs Blender 4.5 and its body extension. See the runtime policy
and three installation levels for details.
1. **Clone and choose an installation level.** The installer asks whether to
 prepare Static, Human reconstruction or Robotics motion generation before
 it changes an environment.
 [code] 
 For a noninteractive run, answer with `--level static`, `--level human` or
 `--level robotics`; no level is chosen by default. `--plan` shows the
 selected route without installing.
2. **Finish the selected level's model and Blender setup.** Static needs
 Pi3X/SAM3 checkpoints and Blender.
 Level 2 also needs PMPose, GVHMR and body skinning,
 installed in their separate environments. Level 3 adds
 Kimodo, native G1 and text models to Static.
 The selector prepares Level 2's Static and SAM3 decoding code, then names
 its remaining installation steps.
3. **Check the selected route and start a scene.** For native G1 motion:
 [code] 
 To reconstruct your own video, put it under `references/`, describe the
 scene with `bash tools/indoor intake`, and follow the
 scene workflow. The examples
 cover generated motion after its model and runtime are configured.
Browse the assets with `python tools/asset_index.py tree`. Setup
covers a source-only install and troubleshooting.
## Tools
### Scene pipeline: `tools/indoor`
A single command drives scene work from request to delivery. Runs are resumable,
and each one gets its own directory under `runs/`.
| Command | What it does |
| --- | --- |
| `intake` | Turn a scene request into a plan and list any missing decisions |
| `scenes`, `plan` | List scenes; summarize a recipe's stages and runtime without running anything |
| `preflight` | Check the recipe, inputs and optional segmentation bounds before a run |
| `run`, `batch` | Run one recipe in the foreground, or queue several in a background worker |
| `status` | Show run and worker progress |
| `preview`, `review-layout`, `accept-preview` | Render a full-length preview and record reviews |
| `check-placement` | Check saved-ro