Plann3r
Predicting planning costs grounded in 3D — CoRL 2026
Plann3r asks a VGGT-style 3D foundation model to do more than reconstruct geometry: to say how expensive it would be to get somewhere. We fine-tune VGGT with a new prediction head that maps a single RGB image, conditioned on a goal, to a geodesic costmap — so planning costs come out grounded in the model’s own 3D representation rather than bolted on by a separate planner.
My contributions:
- Designed and trained the costmap head on top of the fine-tuned VGGT backbone.
- Built the evaluation on HM3D scenes.
- Ran the rebuttal experiments: a loop-closure extension that relaxes the single-traversal assumption, retrieval-noise sweeps, and bootstrap confidence intervals.
Accepted to the Conference on Robot Learning (CoRL) 2026 as first author (missing reference); presenting in person in Austin, TX in November 2026. A full public release is coming soon.