AWSM: Agentic World Simulation and Mapping

AWSM (“awesome”) studies how large-model agents can turn visual observations into editable 3D scenes grounded by camera-pose, depth, and visual–inertial constraints. The package combines a bilingual interactive research article, frozen reconstruction assets, evaluation data, provenance records, and a map-based multi-robot simulation demonstration.

• Interactive article
• Source repository
• Author: Phygital AI
• Published: October 1, 2026

Main result

In one controlled NVIDIA Isaac Sim scene, four complete reconstruction routes are compared over the same 180 RGB identities and editable-output objective:

• M1: RGB-only modeling with GT-assisted Sim(3) alignment for diagnostic evaluation
• M2: RGB-only ViPE poses with pose-conditioned Depth Anything 3 observations
• M3: monocular-inertial ORB-SLAM3 poses with the same class of depth observations
• M4: ground-truth camera poses as a diagnostic route

From M1 to M4, bidirectional surface error decreases from 0.376 m to 0.072 m (about 81%), and final-scene depth AbsRel decreases from 16.44% to 7.60% (about 54%). These are complete-pipeline comparisons, not an isolated IMU ablation.

Repository contents

• index.html, en.html, zh.html: complete bilingual research article
• data/: frozen tables, fixed views, source contracts, citations, and demo metadata
• evidence/: route manifests, SHA256 checksums, publication manifest, audits, and provenance notes
• models/M1–models/M4: frozen Blender and GLB reconstruction assets
• assets/: figures, fixed-view renders, Office Café case-study geometry/posters, and GT comparison geometry
• scripts/: publication build and validation utilities

The canonical MP4 media remain available from the interactive article and source repository. They are not mirrored in this Hub snapshot because repeated transfers to the Hub's large-file endpoint were interrupted. All article text, numerical data, manifests, checksums, frozen Blender/GLB scenes, and still-image evidence are included.

Intended uses

• Reproduce and inspect the reported reconstruction comparisons
• Study editable object-centric Real2Sim outputs and their failure modes
• Reuse frozen scenes, manifests, and evaluation subsets for research
• Inspect the map-based multi-robot execution example

Limitations

• The controlled comparison covers one synthetic scene and one engineering run per route.
• The routes are complete systems, not a strict single-variable ablation.
• M4 uses GT poses and is diagnostic, not deployable.
• M1 metric-space results use GT-assisted Sim(3) alignment.
• The 100-view set uses RGB observations available during modeling and is not a held-out novel-view benchmark.
• The robot demonstration uses predefined routes and simulator-pose feedback; it is not a language-to-plan or online-localization benchmark.
• Persistent memory, map maintenance, and sim-to-real policy improvement remain research directions.

See evidence/provenance-note.md and evidence/publication_manifest.json before reusing or interpreting historical downstream records.

License and paper status

No license has been declared for the source publication package. This private repository does not grant reuse rights beyond those provided by applicable law or separate permission from the rights holder.

AWSM does not currently have an arXiv identifier, so it is not indexed as a Hugging Face Paper. The canonical citation is the interactive research article:

@misc{awsm_2026,
  title        = {{AWSM}: Agentic World Simulation and Mapping},
  author       = {{Phygital AI}},
  year         = {2026},
  howpublished = {Interactive research article},
  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
}