CorrectNav: Self-Correction Flywheel Empowers Vision-Language-Action Navigation Model

PKU-Agibot Lab *Equal contribution, † Project Leader, ‡ Corresponding author https://correctnav.github.io Crowded Objects Avoidance Open-vocabulary Landmark Move Forward and Turn right at the human-like robot. Continue moving to stop near the yellow box. … Z-Shape Building Structure Walk down the corridor hallway in front of you and you will see an opened meeting room. Enter the … … … Walk straight along the hallway until you reach the red fire extinguisher box at the end and stop when you reach … Pedestrian Avoidance … Error Correction … in front of a white wall, turn right. Walk forward. When you see a green plant on your right front, stop. Drift Correction Walk straight and turn left in front of a wall. Walk straight and turn right at the opened door. Enter and walk to the wooden table. … … Walk until you reach the plant and turn left. Walk straight, turn left at the next corner, walk forward to the … Instruction Across Rooms Walk out of the kitchen room you are in and turn left. Move across the living room, walk to the end of the hallway and turn right .Walk into the bedroom and stop by the bed. Landmark State Change Move forward and turn right to walk through an opened doorway. … … … … … … C rrectNa Figure 1: Diverse Capabilities of CorrectNav. The model takes only monocular RGB video and language instructions as inputs, predicting navigation actions. Empowered by the Self-correction Flywheel post-training, CorrectNav not only maintains outstanding multimodal reasoning (Blue), but also displays improved deviation correction (Red), obstacle avoidance (Green), and complex action execution (Yellow).