# OneCanvas: robot AI research model

OneCanvas (3D Scene Understanding via Panoramic Reprojection) is a 2026 research method from [Technical University of Munich](/companies/tum-robotics.md) and [Huawei](/companies/huawei.md) (Baranowski, Chen, Nießner; NeurIPS 2026 Spotlight) that lifts multi-view RGB-D patch features onto one equirectangular panoramic canvas so a frozen 2D VLM can do 3D scene understanding and situated reasoning with far less training compute than geometry-encoder or large-data alternatives.

- **Slug:** onecanvas
- **Type:** Research model
- **Maturity (DEPLOY ladder):** Research
- **Open source:** no

## Architecture

Frozen VLM vision encoder (Qwen3-VL-8B class) extracts per-view patches; depth + pose lift to 3D; continuous equirectangular canvas tokens with native MRoPE (longitude/latitude/frame) plus additive metric 3D position embedding; LoRA adapters with two-stage spatial curriculum then target QA adaptation.


## Key facts

- **SQA3D 65.3 EM@1 / VSI-Bench 71.3 / SPBench 72.1 zero-shot (research-stated):** Research-stated (research-stated; also arXiv; Robots Digest X): OneCanvas (Baranowski / Chen / Nießner; TUM + [Huawei](/companies/huawei.md); NeurIPS 2026 Spotlight) reports SOTA spatial reasoning on SQA3D (65.3 EM@1, +2.3 over prior best), VSI-Bench (71.3 average), and SPBench (72.1 zero-shot overall, +4.8), claiming an order of magnitude less training compute than closest competitors. Cap: author project-page / paper tables; not an independent replication.
- **Panoramic patch-feature canvas for frozen 2D VLMs (research-stated):** Research-stated (research-stated; also arXiv): method lifts multi-view patch features into 3D via depth and camera pose, places each patch as a continuous token on one equirectangular panorama (longitude/latitude from a chosen origin), adds a metric 3D position embedding, and feeds a pretrained VLM (Qwen3-VL-8B class) with LoRA adapters only. Cap: author method description.


## Sources (3)

1. **OneCanvas: 3D Scene Understanding via Panoramic Reprojection** · https://baranowskibrt.github.io/onecanvas/ · 2026
2. **OneCanvas: 3D Scene Understanding via Panoramic Reprojection (arXiv:2606.19253)** · https://arxiv.org/abs/2606.19253 · 2026
3. **Robots Digest: OneCanvas summary** · https://x.com/robotsdigest/status/2106510045513154969 · 2026-10-03


## Common questions

### What is OneCanvas?

OneCanvas (3D Scene Understanding via Panoramic Reprojection) is a 2026 research method from [Technical University of Munich](/companies/tum-robotics.md) and [Huawei](/companies/huawei.md) (Baranowski, Chen, Nießner; NeurIPS 2026 Spotlight) that lifts multi-view RGB-D patch features onto one equirectangular panoramic canvas so a frozen 2D VLM can do 3D scene understanding and situated reasoning with far less training compute than geometry-encoder or large-data alternatives.

### Is OneCanvas open source?

No. OneCanvas is recorded as proprietary on the DEPLOY registry. Model weights and source are not publicly available.

### What type of AI is OneCanvas?

OneCanvas is a research model, built on a Frozen VLM vision encoder (Qwen3-VL-8B class) extracts per-view patches; depth + pose lift to 3D; continuous equirectangular canvas tokens with native MRoPE (longitude/latitude/frame) plus additive metric 3D position embedding; LoRA adapters with two-stage spatial curriculum then target QA adaptation. architecture on the DEPLOY registry.

### What is OneCanvas's maturity stage?

OneCanvas is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file.

### Which robots run on OneCanvas?

No robot models on the DEPLOY registry are recorded as running OneCanvas. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.


_API: GET /v1/brains/3fb71384-65aa-44f5-ae29-d406a68ffedf · canonical URL: /brains/onecanvas_