ubr-physical-ai/ubr-maze-nav
ubr-maze-nav — vision + instrument waypoint planning for a small tracked robot Synthetic navigation corpus for fine-tuning small vision-language models to plan local waypoint paths for a 0.3 m tracked ground robot in corridor/maze environments, plus the frozen evaluation suite used in our internal reports. Each sample is one first-person RGB frame (640×480) from the robot's camera in a procedurally generated MuJoCo scene, an instruction carrying the goal (bearing/range) and a… See the full description on the dataset page: https://huggingface.co/datasets/ubr-physical-ai/ubr-maze-nav.
ubr-maze-nav — vision + instrument waypoint planning for a small tracked robot
Synthetic navigation corpus for fine-tuning small vision-language models to plan local waypoint paths for a 0.3 m tracked ground robot in corridor/maze environments, plus the frozen evaluation suite used in our internal reports.
Each sample is one first-person RGB frame (640×480) from the robot's camera in a procedurally generated MuJoCo scene, an instruction carrying the goal (bearing/range) and a short instrument channel (prose ranging readings, as the deployed sensor stack reports them), and a completion of ego-frame waypoints (x, y, heading) derived from an omniscient expert planner. Corridor scale, robot pose, lighting, materials, and camera are randomized; 30 % of frames keep a fixed flat rendering. Collision-checked labels only (a validity gate rejects frames whose expert path violates the robot's clearance disc).
Contents
The full corpus lineage is included so every training-data ablation is reproducible. Two pairs share byte-identical images and differ only in the instruction text — they isolate the instrument-channel format: corpus_v1_r/corpus_v1_ri and corpus_v2_fan/corpus_v2_prose.
data.jsonl per split; images referenced relatively. Rows carry the scene seed and exact robot pose, which is how our collision metric reconstructs the true wall grid at eval time.
Intended use & scope
Built for research on instrument-conditioned local planning with compact VLMs (≈4B) on edge hardware. Scenes are kinematic and synthetic; this corpus makes no sim-to-real claim by itself. The generator, training recipes, and result reports are not part of this artifact.
Contact
UB Robotics / teex — via the organization page.
