Qwen unveils one navigation model for five robot tasks
Qwen's RobotNav model uses one set of weights for five navigation domains and lets an agent adjust its visual context for each call, which makes it a reusable building block for agentic robot systems.
Qwen has introduced Qwen-RobotNav. The navigation model is built on Qwen3-VL and handles instruction following, object search, target tracking, embodied question answering and autonomous driving with a single set of weights.
The main design idea is what Qwen calls a controllable observation protocol. The company compares it to the Model Context Protocol for tool use. Four settings are exposed at inference time: visual token budget, temporal decay, per-camera weighting, and frame sampling mode (random or latest). All four were randomized for each sample during training, so Qwen says the model works with any configuration without retraining. The architecture adds a 4-layer MLP action head to the backbone. The head outputs eight waypoints with position and heading, and camera identity and time order are marked with natural-language tags.
The model was trained on 15.6 million samples. Qwen reports state-of-the-art results across the five domains:
- 76.5% success rate on VLN-CE RxR and 72.1% on R2R (8B model)
- 75.6% success rate on HM3Dv2 object-goal navigation using RGB only, which beats depth-based methods
- 90.0% tracking rate on EVT-Bench
- 91.4 PDMS on NAVSIM
- New best results on three EQA benchmarks
Performance improved steadily from 2B to 8B parameters.
In a two-tier setup, Qwen3.7-Plus acts as the planner. It breaks long goals into sub-goals and sends navigation calls that set the task mode and observation settings. Compared with the previous best, this setup scored 15.4% higher on EXPRESS-Bench and used 77% fewer navigation steps.
Qwen also says it ran the model zero-shot on a Unitree Go2 quadruped using the robot's built-in low-resolution camera. The post does not say whether the weights or code will be released.