From going viral on GitHub to earning an ECCV Oral, developers worldwide have pushed LingBot-Map to new heights.

Robots are learning a capability that was once exclusive to humans: remembering the world they have traversed.

When entering an unfamiliar room for the first time, humans do not store everything they see frame by frame. We remember key structures: where the door is, where the table is, the positions we just passed, and where we are currently located.

It is this selective memory that supports humans to act continuously in complex environments. For robots, this is also a problem that must be solved to move into the real world.

In April this year, the streaming 3D reconstruction foundation model LingBot-Map open-sourced by Ant Lingbo explored a new technical route focusing on the core issue of "how robots remember spaces". Without complex hardware, with only an ordinary RGB camera, robots can complete camera pose estimation and 3D scene structure reconstruction in real time during video collection, thus filling the key technical gap in the field of real-time spatial perception.

In long sequences traversing multiple rooms, LingBot-Map still demonstrates extremely strong robustness in the face of drastic environmental changes and large viewport shifts

As of now, the official GitHub project of LingBot-Map has obtained 16.9K Stars and 1.9K Forks, and has been on GitHub Trending for many times, becoming one of the most notable open-source projects in the 3D vision field.

• Project address: https://github.com/Robbyant/lingbot-map
• Paper address: https://arxiv.org/pdf/2604.14141
• Paper homepage: https://technology.robbyant.com/lingbot-map
• Hugging Face link: https://huggingface.co/robbyant/lingbot-map
• ModelScope link: https://www.modelscope.cn/models/Robbyant/lingbot-map

Surprisingly, this research has now received new recognition.

Its paper "Geometric Context Transformer for Streaming 3D Reconstruction" has been selected as an ECCV 2026 Oral and will be presented orally during the conference. At the ECCV opening ceremony yesterday, the on-site editor from Jiqizhixin also spotted this work on the Award Candidates list.

From high attention in the GitHub community to recognition from top vision conferences, LingBot-Map has completed dual validation from the developer ecosystem to the academic world.

More importantly, it reflects a change: the competition for robot spatial perception is shifting from seeing the world to remembering the world.

Prior to Top Conference Recognition, LingBot-Map Has Completed Open-Source Community Validation

Before being selected for ECCV Oral, LingBot-Map had already gained massive attention in the open-source community, with developers carrying out various secondary developments around the project.

Some quickly completed reproduction and verification, some ported it to different hardware environments for trial operation, and other developers began to explore new spatial data collection methods.

Everyone's usage is very interesting, and this active exploration from the community often better reflects the vitality of the project than the mere number of Stars.

Enabling more people to experience quickly: from deployment to visualization

For many research projects, the publication of papers does not mean that ordinary developers can actually use them. Environment configuration, dependency installation, result display and other links often become the gap between the paper Demo and actual experience.

Shortly after LingBot-Map was open-sourced, a developer produced a native desktop front-end and 3D point cloud viewer that can run locally on Apple Silicon Mac, allowing users to more intuitively view the spatial reconstruction results generated by the model.

Source: https://github.com/donalleniii/lingbot-desktop-mac

At the same time, some developers produced a one-click startup solution based on Pinokio, which further encapsulates the environment configuration, dependency installation and Demo operation process, enabling users without complex configuration experience to quickly experience LingBot-Map.

Source: https://github.com/cocktailpeanut/lingbot-map.pinokio

These attempts fill the important gap between research results and development experience, allowing more users to actually access and use this streaming 3D reconstruction model.

Operation on more devices: consumer-grade hardware adaptation

In addition to optimizing the usage process, developers have also begun to try to expand the hardware compatibility scope of LingBot-Map.

Developer ureeey carried out adaptation attempts for devices with RTX 4060 8GB VRAM, reducing VRAM pressure by optimizing operation strategies, so that LingBot-Map can run in more lightweight GPU environments.

Source: https://github.com/ureeey/lingbot-map-rtx4060-8g/blob/main/README.md?utm_source=chatgpt.com

The significance behind such attempts is: when the model is no longer limited to high-configuration workstations and can be accessed in more ordinary development environments, the technology will have the opportunity to be tested, modified and applied by more people.

From cameras to smart glasses: exploring new collection entrances

On the X platform, some developers noted that in the past, some high-precision 3D scanning tasks usually relied on professional LiDAR devices or required complex offline optimization processes, while traditional methods also easily faced stability challenges under long-term continuous input.

In contrast, LingBot-Map achieves real-time streaming reconstruction through ordinary video streams, runs with only a single GPU at about 20 FPS, and can process long sequence inputs of more than 10,000 frames.

Source: https://x.com/XAMTO_AI/status/2080900857235726635

In the view of these developers, LingBot-Map demonstrates a more lightweight spatial perception path: reducing the dependence on dedicated hardware and complex processes, enabling real-time 3D reconstruction capabilities to enter more scenarios such as robotics, AR/VR, and autonomous driving.

Based on this, some people have further connected LingBot-Map to consumer-grade devices.

He modified Ray-Ban Meta Gen-2 smart glasses (no LiDAR required) into a 3D scanning device, collected data through a mobile App, and called the LingBot-Map backend deployed on RunPod to complete 3D mapping of the space in about 30 seconds. Even with fewer video frames, the indoor scene can still obtain good reconstruction results.

Source: https://x.com/0x6rss/status/2079597540568137866

In addition to smart glasses, this user also ran LingBot-Map on an ordinary Dell desktop, and obtained an interactively browsable 3D point cloud model after about 61.5 seconds of video collection only aimed at the desk.

Source: https://x.com/shavonnewong_/status/2080130333094101360

These are only part of the community explorations after LingBot-Map was open-sourced. For a 3D vision research project, such continuous secondary development is uncommon, which indicates that LingBot-Map has begun to enter real usage scenarios.

After open-sourcing, the team has also been continuously updating evaluation scripts, inference optimizations, example cases and issue fixes to further improve the usability and stability of the project. At the same time, the official also demonstrated the ultra-long video reconstruction result: in an indoor walking video of about 25,000 frames lasting 13 minutes, LingBot-Map can still maintain continuous reconstruction and generate a complete 3D spatial representation.

Of course, community popularity is only one aspect. For a research result, what truly determines its long-term influence is still whether it solves the key problems in the field.

Behind ECCV Oral, How Does LingBot-Map Redesign Robot Spatial Memory?

So what exactly problems has LingBot-Map solved? Why can a streaming 3D reconstruction model attract continuous expansion from developers and gain recognition as ECCV Oral?

The answer goes back to the most core challenge in robot spatial perception: how to enable robots to maintain long-term memory of space while avoiding infinite increase in computing and storage costs over time in the face of growing video input.

For robots, 3D reconstruction requires understanding the environment while moving. When a robot sees the current frame, it needs to answer two questions at the same time: where is the position it is seeing now? What is the relationship between the areas it has passed in the past and the current frame?

This means the model needs to continuously estimate the camera trajectory, generate corresponding depth information, and finally form an expanding 3D map.

Compared with offline reconstruction, streaming reconstruction faces a core contradiction: retaining more historical information can improve spatial consistency, but as the history gets longer, the computing and storage costs will also grow rapidly.

Existing methods generally adopt three routes.

One method is similar to remembering all history. For example, schemes based on causal attention will cache the information of all previous frames. Such a model has complete context, but as the sequence grows, the VRAM and computation amount also keep increasing.

Another method chooses to compress the historical state, such as a recurrent network structure. However, excessive compression may cause the model to forget key geometric information, leading to trajectory drift after long-time operation.

There are also some methods that combine traditional SLAM to maintain stability through key frame selection and optimization algorithms. However, such methods rely on manually designed rules and require additional optimization processes, making it difficult to balance real-time performance.

Therefore, the really difficult problem is: in the face of a growing video stream, which spatial information must the robot retain, and which information can be discarded?

LingBot-Map refers to this problem as context management in streaming reconstruction. It hopes to make the model retain only the most important geometric clues like human spatial memory, instead of recording all observations.

GCA: Dividing Spatial Memory into Three Layers to Make the Model Know What to Remember

To solve this problem, LingBot-Map proposes Geometric Context Attention (GCA).

Its core idea comes from traditional SLAM: a stable spatial system usually requires three types of information: a fixed reference point to determine coordinates, nearby area information to judge the current position, and long-distance history to avoid cumulative drift. Therefore, GCA splits the streaming state into three parts.

LingBot-Map workflow. The framework processes the current view relative to the initialization set [T, T). The DINO backbone extracts image features, which are then refined through alternating frame attention layers and GCA layers. In the GCA module, the input view aggregates information from the anchor context, the local pose reference window [T, T], and the trajectory memory context. Finally, the task-specific head predicts the camera pose and depth map, enabling robust, memory-efficient streaming 3D reconstruction for long sequences.

First layer: Anchor Context, establishing spatial coordinates. It is responsible for remembering "where I started". It provides a stable benchmark for the entire 3D coordinate system. Spatial reconstruction is most afraid of coordinate drift. With anchors, the model still clearly knows where the first frame was located when processing the 10,000th frame.

Second layer: Local Pose-Reference Window, maintaining local precise positioning. It is responsible for capturing local geometric details near the current position. This is equivalent to maintaining clear real-time perception of "what is around me", ensuring the accuracy of frame-by-frame reconstruction.

Third layer: Trajectory Memory, compressing long-term history to avoid drift. This is a relatively critical design in the entire architecture. It compresses huge historical information into extremely compact frame-by-frame tokens, retaining the "impression" of past paths at a low storage cost. It is this mechanism that makes the memory consumption of LingBot-Map hardly grow with the video length, and the total computation amount and memory usage remain almost at the same level when processing 100 frames and 10,000 frames.

The efficiency improvement brought by this design is quite considerable. In a 10,000-frame sequence, traditional causal attention requires about 5 million tokens; GCA only needs about 70,000 tokens, which is a full compression of nearly 80 times, and when processing each new frame, the computation amount and memory consumption hardly grow with the total number of frames.

Attention mask comparison. Each box represents a token of a frame, consisting of a small segment of context tokens and a larger segment of image tokens. (a) Full attention focuses on all frames. (b) Causal attention supports streaming processing, but the computational overhead grows linearly with the sequence length. (c) Sliding-window attention limits the computational cost but loses long-range context. (d) GCA divides the streaming context into anchors (n=2), local window (k=2) and trajectory memory, which keeps the computational cost almost constant as the sequence length increases while retaining rich long-range context information.

How does it perform on benchmark tests