Turn a mobile chassis into a robot that actually gets work done. Self-developed 3D multi-line LiDAR SLAM, multi-sensor fusion, dynamic obstacle avoidance and embodied decision-making — one stack driving 7 chassis structures, cleaning robots, composite robots and dual-arm wheeled humanoids.
We do not build a generic big-data platform. This page describes the four-layer algorithm and software stack that lets a robot see, localize, move and understand — fully self-developed, trimmable per robot form, and reused across several product lines.
3D multi-line LiDAR as the primary sensor, optionally with front/rear dual LiDAR, depth cameras and ultrasonic radar for 360° stereo perception without blind spots; point-cloud pre-processing, ground segmentation and multi-target tracking output a stable traversable area.
LiDAR-inertial fused 3D SLAM builds a 3D point-cloud map in real time. No reflective plates, magnetic tape or QR codes required; 10 mm re-localization accuracy and continuous navigation across large sites and multiple floors.
Global path planning combined with local dynamic avoidance lets the robot pass under overhanging objects and through narrow aisles and low obstacles; automatic elevator riding and auto-docking enable end-to-end unmanned operation.
Integration with DeepSeek, QWQ, ERNIE Bot, Spark and ChatGPT. Voice commands are semantically parsed and decomposed into tasks, then handed to the navigation layer — so the robot understands human phrasing such as "go to meeting room 3".
Sensors to actuators form a single one-way pipeline; data logging and OTA run as a side channel. Each stage can be trimmed or replaced without affecting the others.
| Item | Conventional marker-based AGV | Tbot 3D multi-line SLAM |
|---|---|---|
| Site markers | Reflective plates, magnetic tape or QR codes required | None — map and go |
| Perception | 2D single-line, same-height obstacles only | 3D multi-line, includes overhanging and low obstacles |
| Re-localization accuracy | Depends on marker layout | 10 mm |
| Chassis endurance | — | ≥10 h unloaded · 6 h loaded |
| Multi-robot coordination | External fleet manager required | Group scheduling and remote dashboard supported |
Native ROS / rosbridge WebSocket interfaces with complete Topic / Service definitions from chassis to application, plus examples in Python, JS, C++, C, Java, C# and Go. The same stack already runs on 7 chassis structures — N5 / OW / S / X / X2 / ACK / MW. Changing the form does not change the algorithm, which is why ODM delivery can be fast.
| Layer | What you can do | How |
|---|---|---|
| Application | Change business logic, connect your own fleet manager | ROS Topic / Service, WebSocket |
| Parameters | Speed, avoidance radius, path smoothing and similar | Config push, no code change |
| Sensors | Add depth camera, ultrasonic, second LiDAR | Connect via interface and re-calibrate |
Standard projects follow the sequence below. Actual timing depends on site complexity and customization scope — 15–25 days for standard models, 45–60 days for custom ODM.
No single SLAM stack fits every environment. Being explicit about what it does and does not handle is more useful than staying silent.
Related termsSLAMOccupancy GridSensor FusionPath Planning
One-stop customization of appearance, structure, hardware, software and platform — from prototype validation to mass production.