# TrackEverything: robot AI research model

TrackEverything is a Sep 2026 academic 3D point tracker (Jain, Paruchuri, Gupta, Zhang, Schmidt, Engel, Fragkiadaki, Harley; arXiv 2609.30222) that represents videos as persistent 3D scene tracks in world coordinates. Voxel de-duplication at sliding-window boundaries merges co-located tracks so memory scales with unique scene geometry rather than clip length; 3D WAFT samples features in the scene cloud instead of building 4D correlation volumes. Research demo for long-horizon dense tracking / robotics / spatial / 4D reconstruction; not a commercial product.

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

## Architecture

Persistent 3D scene tracks in world coordinates; voxelization-based de-duplication at sliding-window boundaries; endpoint refiner with static-vs-dynamic classification; lightweight trajectory refiner for dynamic points; 3D WAFT feature sampling in the scene cloud


## Key facts

- **System (project-stated):** Project-stated (project-stated): TrackEverything (Jain / Paruchuri / Gupta / Zhang / Schmidt / Engel / Fragkiadaki / Harley; arXiv 2609.30222) is a 3D point tracker that represents videos as persistent 3D scene tracks in world coordinates so complexity scales with unique scene geometry rather than video length. Academic research system, not a commercial product
- **Voxel de-duplication + 3D WAFT (project-stated):** Project-stated (project-stated): voxelization-based de-duplication at sliding-window boundaries merges co-located tracks; tracking decomposes into endpoint refinement plus static-vs-dynamic classification and a lightweight trajectory refiner for dynamic points; 3D WAFT replaces memory-heavy 4D correlation volumes with feature sampling in the scene cloud. Cap: paper-reported method
- **TAPVid-3D dense-tracking margin:** The paper reports that on TAPVid-3D, TrackEverything outperforms open-source all-frame dense 3D trackers by more than 20% APD on short clips while tracking all visible points (arXiv 2609.30222).
- **Long-horizon memory bound:** The authors say TrackEverything is, to their knowledge, the first 3D tracker that can track all visible points across videos longer than 1000 frames within 40 GB of GPU memory (arXiv 2609.30222).
- **CMU and Meta authors:** The paper lists authors affiliated with [Carnegie Mellon University](/companies/carnegie-mellon-university.md) and [Meta](/companies/meta.md) and points to the project site trackeverything.github.io (arXiv 2609.30222).


## Sources (3)

1. **TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations (arXiv:2609.30222)** · https://arxiv.org/abs/2609.30222 · 2026-09-30
2. **TrackEverything PDF (arXiv:2609.30222)** · https://arxiv.org/pdf/2609.30222 · 2026-09-30
3. **Secondary summary: TrackEverything voxel de-duplication (robotsdigest)** · https://x.com/robotsdigest/status/2105304160451056084 · 2026-09-30


## Common questions

### What is TrackEverything?

TrackEverything is a Sep 2026 academic 3D point tracker (Jain, Paruchuri, Gupta, Zhang, Schmidt, Engel, Fragkiadaki, Harley; arXiv 2609.30222) that represents videos as persistent 3D scene tracks in world coordinates. Voxel de-duplication at sliding-window boundaries merges co-located tracks so memory scales with unique scene geometry rather than clip length; 3D WAFT samples features in the scene cloud instead of building 4D correlation volumes. Research demo for long-horizon dense tracking / robotics / spatial / 4D reconstruction; not a commercial product.

### Is TrackEverything open source?

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

### What type of AI is TrackEverything?

TrackEverything is a research model, built on a Persistent 3D scene tracks in world coordinates; voxelization-based de-duplication at sliding-window boundaries; endpoint refiner with static-vs-dynamic classification; lightweight trajectory refiner for dynamic points; 3D WAFT feature sampling in the scene cloud architecture on the DEPLOY registry.

### What is TrackEverything's maturity stage?

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

### Which robots run on TrackEverything?

No robot models on the DEPLOY registry are recorded as running TrackEverything. 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/4888ee4c-71f5-4e69-8b49-030297ec0c16 · canonical URL: /brains/trackeverything_