LightNav-0 Released

Light Origins has released LightNav-0, which directly leverages the spatial reasoning capabilities of a pre-trained Vision-Language Model (VLM) for robot control. Using Qwen3-VL-4B-Instruct as the backbone, the model performs instruction following, object finding, and target tracking based solely on natural language instructions, without dedicated navigation modules or task-separation tokens.

Core Technology and Architecture

• Single-Model Integration: Outputs spatial reasoning and action codes through the backbone's autoregressive language model head without separate heads. Action outputs are decoded into 3 Residual Vector Quantization (RVQ) action tokens to generate 10-step SE(2) waypoints.
• Minimized Sensor Dependency: Uses only a single front-facing RGB stream, requiring no depth sensors, odometry, or panoramic equipment.
• Robot Transferability: Uses a common geometric interface independent of the robot body, allowing transfer to various robots such as wheeled, quadruped, and humanoid types with the same checkpoint without retraining.

Performance and Benchmark Results

• Instruction Following (VLN-CE): Achieved SR 68.5% and SPL 62.8% on R2R val-unseen, demonstrating superior performance compared to existing monocular systems.
• Embodied Visual Tracking (EVT-Bench): Achieved the highest success rates with SR 91.7% in Single Target Tracking (STT) and SR 82.6% in Distracted Tracking (DT). It was reported to outperform panoramic/multi-camera systems, particularly in environments with distractors.
• Scaling Analysis: Performance improved when scaling the backbone from 2B to 4B parameters, but degraded at 8B. Diversity in the training environment contributed more to performance improvement than the number of parameters.

Deployment and Usage

• Released under the Apache 2.0 license, providing simulation benchmark harnesses and deployment code.
• The architecture runs the model on a GPU host while the robot operates as a thin WebSocket client, including a ROS 2 stack and adapters (such as Unitree Go2) for integration with low-level controllers during real robot deployment.