🐹 3D HAMSTER — 3D Trajectory Prediction

Depth-aware Vision-Language-Action planner (Qwen3-VL-8B + LingBot-Depth geometry encoder). From a single RGB image + metric depth map + language instruction, it predicts a metric 3D end-effector trajectory ([u, v, depth] waypoints + gripper states), plus 2D trajectory / pointing / bounding-box / VQA modes.

Paper · Model · Code · Project page

⚠️ Depth must be metric (meters) and aligned to the RGB frame. Disparity / normalized / millimeter depth will degrade the geometry.