mermaidliu/andrea-nextwin — explained in plain English
Analysis updated 2026-05-18
Run a demo where a Unitree G1 humanoid robot performs a simulated rescue task from a natural language instruction.
Prototype a perception pipeline combining lidar point clouds and camera object detection for a humanoid robot.
Visualize a robot's sensor data and planned actions in a 3D digital twin in the browser.
Experiment with connecting an LLM, or a fallback rule based parser, to turn instructions into a robot action blueprint.
| mermaidliu/andrea-nextwin | adityasharmadotai-hash/docs-reader-rag-agent | alekseiul/hermes-researcher-agent | |
|---|---|---|---|
| Stars | 29 | 29 | 29 |
| Language | Python | Python | Python |
| Setup difficulty | hard | easy | moderate |
| Complexity | 5/5 | 2/5 | 2/5 |
| Audience | researcher | vibe coder | researcher |
Figures from each repo's GitHub metadata at analysis time.
Full functionality needs a real Unitree G1 robot, Livox lidar, RealSense camera, and ROS2 setup, the UI-only demo installs with far fewer dependencies.
Andrea-NexTwin, also called NexTwin, is a research and demo project that builds a physical world model and language driven control system for humanoid robots. The stated goal is to let a robot see and understand its surroundings, understand a task described in natural language, and then act on that task in real physical space, using the Unitree G1 humanoid robot as the target hardware. The system is organized into five layers that can be developed somewhat independently: a sensor layer that reads data from the robot's Livox lidar and RealSense camera over ROS2, a vision layer that splits the camera feed into four directional views, a perception layer that runs YOLO object detection on those views, a fusion layer that combines the lidar bird's eye view with the detection results and a rule engine to decide what to do, and a control layer that is meant to turn those decisions into actual robot movement such as turning, walking forward, or pushing an object. The control layer is currently a mock stand in that a developer is expected to connect to the real robot. A bundled demo scenario shows the G1 robot receiving a rescue instruction, scanning the area with its sensors, identifying a smaller robot pinned under debris using YOLO, planning a response with the rule engine, and acting on it, with the whole sequence shown on a 3D digital twin built with Three.js. The project can optionally connect to an OpenAI compatible large language model to parse natural language instructions, but falls back to a built in rule based parser if no API key is provided, so the demo works without any AI service configured. The backend is written in Python with FastAPI, and the frontend is plain HTML, CSS, and JavaScript with Three.js for the 3D scene. Beyond the rescue demo, the project also includes a user workspace for configuring scenes and workflows, a developer center for uploading SDKs and world models, and an API platform page describing subscription tiers and per call billing, though these read as product mockups rather than a live commercial service. Setup starts with a lightweight install that only pulls in a handful of packages to run the web interface, with heavier dependencies like OpenCV and YOLO installed separately only if needed.
A demo system that gives a humanoid robot a physical world model, letting it interpret natural language tasks, perceive its surroundings with lidar and cameras, and act in real space.
Mainly Python. The stack also includes Python, FastAPI, ROS2.
Permissive MIT license, free to use, modify, and distribute including commercially, as long as you keep the copyright notice.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.