Plann3r checkpoints

Checkpoints for "Plann3r: Predicting Planning Costs Grounded in 3D" (CoRL 2026). Plann3r predicts a 16x16 geodesic costmap over a query image from eight map images and a goal pixel, and a GNM controller turns the costmap into waypoints.

Code: https://github.com/MostlyKIGuess/plann3r-code Project page: https://plann3r.github.io/

Files

Path	Contents
planner/checkpoint_best.pt	Plann3r planner used for the reported results
planner/ablations/costmap_only.pt	ablation, costmap loss only
planner/ablations/no_grad_loss.pt	ablation, without the gradient loss
planner/ablations/no_pointmap_loss.pt	ablation, without the pointmap loss
planner/ablations/frozen_mlp_goal_token.pt	ablation, frozen backbone
controller/predicted_costmap/latest.pth	controller trained on Plann3r costmaps, used for the reported results
controller/gt_trained/latest.pth	controller trained on simulator geodesic costmaps

VGGT-1B (facebook/VGGT-1B) and MegaLoc (gberton/MegaLoc) are not included. Download them from their authors.

Usage

export PLANN3R_ROOT=/absolute/path/to/plann3r-release
hf download MostlyK/plann3r --local-dir "$PLANN3R_ROOT/models"

Then follow docs/setup.md and docs/evaluation.md in the code repository.

License

The planners are fine-tuned from VGGT-1B, so the VGGT-1B license applies.

Citation

@inproceedings{vadali2026plann3r,
  title     = {{Plann3r}: Predicting Planning Costs Grounded in {3D}},
  author    = {Vadali, Aditya and Pandya, Krish and Garg, Vansh and Chandna, Nishant and
               Craggs, Dustin and Dayoub, Feras and Krishna, Madhava and Garg, Sourav},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}