what is qwen-robotnav fr?

Qwen-RobotNav is a research project from the Qwen team that builds a single AI model to handle many kinds of robot navigation at once. Instead of training separate models for each task, it combines instruction following, point and object goal navigation, target tracking, autonomous driving, and embodied question answering into one shared model built on top of Qwen3-VL, the team's vision-language model. The main idea is that all these navigation tasks are really about how a robot processes its recent visual history. Following long instructions needs memory of what happened earlier, tracking a moving target needs the freshest high-resolution frames, and driving needs recent motion across multiple cameras. Qwen-RobotNav lets a planner adjust settings like how many visual tokens to use, how much older frames are discounted, how much weight each camera gets, and how frames are sampled, all without retraining the model for each new task. Under the hood, the model takes the Qwen3-VL backbone and adds a small extra network that outputs eight waypoints describing where the robot should go next. It was trained on 15.6 million samples and the README reports it scaling well from a 2 billion parameter version up to 8 billion parameters, plus zero-shot transfer to real robots it was not specifically trained on. The README includes demo videos and benchmark tables comparing Qwen-RobotNav against other navigation models on tasks like instruction following, object search, question answering while exploring, and self-driving car benchmarks, where it performs competitively or better than the compared systems. The team states clearly that there is currently no plan to release the model weights. This repository exists to share technical reports, demos, and benchmark results rather than usable code, so it functions more as a public record of the research than a tool you can install and run.