> For the complete documentation index, see [llms.txt](https://docs.konnex.world/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.konnex.world/supported-ai-models/orb-slam3.md).

# ORB-SLAM3 & RTAB-Map

## ORB-SLAM3 (visual SLAM, widely used)

[ORB-SLAM3](https://github.com/UZ-SLAMLab/ORB_SLAM3) (Campos et al., 2020) is a widely used open-source **visual SLAM** system: it tracks the camera (or camera rig), estimates egomotion, and builds a **sparse 3D map** from ORB feature tracks. It supports **monocular**, **stereo**, and **RGB-D** input, and includes **loop closing** and **multi-map** handling.

**Why it’s relevant for Konnex:** the [SLAM 3D map](/subnets-workload-classes/slam-3d-map.md) subnet is about **mesh / geometry / semantics** and **verifiable** sensor data. A miner can run ORB-SLAM3 (or a derivative) to produce **trajectories**, **keyframe poses**, and **3D landmarks** for validators to compare against ground truth or scoring rules, alongside the [Proof-of-Physical-Work](/understand-konnex/contracts-and-popw.md) bundle.

| Resource | URL                                                                                                                                                |
| -------- | -------------------------------------------------------------------------------------------------------------------------------------------------- |
| Code     | [github.com/UZ-SLAMLab/ORB\_SLAM3](https://github.com/UZ-SLAMLab/ORB_SLAM3)                                                                        |
| Paper    | *ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM* — [arXiv:2007.11898](https://arxiv.org/abs/2007.11898) |

**Outputs:** keyframe poses, **MapPoints** (sparse 3D), trajectory. A **dense** mesh usually needs an extra **fusion** step (TSDF, Poisson, etc.) on top of depth or MVS, depending on your pipeline.

**Hardware:** runs on CPU; GPU use depends on the build. **Visual-inertial** modes expect a calibrated IMU in the stack per upstream documentation.

***

## RTAB-Map (dense 3D maps, popular in ROS)

[RTAB-Map](http://introlab.github.io/rtabmap/) is a common choice in **ROS / ROS2** when teams want **online** mapping with **loop closure** and **export** to occupancy grids, point clouds, or meshes for navigation and inspection. It is often a practical complement to “research” VSLAM stacks: more turnkey for **room-scale** RGB-D / stereo (and many LiDAR–RGB setups) in integration tutorials.

| Resource | URL                                                                |
| -------- | ------------------------------------------------------------------ |
| Project  | [introlab.github.io/rtabmap](http://introlab.github.io/rtabmap/)   |
| Code     | [github.com/introlab/rtabmap](https://github.com/introlab/rtabmap) |

**Picking a stack:** use ORB-SLAM3 as a **reference** VIO/VSLAM baseline; use RTAB-Map when you need **out-of-the-box** mapping exports in a **ROS**-centric 3D mapping workflow.

***

## Konnex alignment

* Miners deliver reconstructions, trajectories, and signed sensor payloads as the subnet contract specifies.
* Validators check fidelity and consistency (geometry, semantics, or both) per subnet design.

## See also

* [SLAM 3D map](/subnets-workload-classes/slam-3d-map.md)
* [PI-0 (manipulation)](/supported-ai-models/pi0.md) — different modality, same “miner–validator” idea
* [AI models overview](/supported-ai-models/ai.md)