SeonghuJeon/3da-libero-gam

GAM LIBERO Checkpoints

This repository contains released GAM (Geometric Action Model) checkpoints for LIBERO and LIBERO-Plus evaluation.

Suite key	LIBERO suite	Checkpoint	Config
spatial	libero_spatial	spatial/gam.pt	spatial/config.yaml
object	libero_object	object/gam.pt	object/config.yaml
goal	libero_goal	goal/gam.pt	goal/config.yaml
long	libero_10	long/gam.pt	long/config.yaml

Pretrained GAM Checkpoint

Name	Checkpoint	Source
pretrained-gam	pretrained/pretrained-gam.pt	SeonghuJeon/da3-shallow12-robocasa25-h8-step235000 (0235000.pt, step 235k)

This checkpoint is provided under the simple pretrained-gam name for convenience. It corresponds to the DA3 shallow12 RoboCasa25 H8 checkpoint at step 235000.

The configs are portable: stage_1.ckpt_path, dataset.hdf5_root, and dataset.stats_dir resolve under ${DA3_ROOT} by default. Download SeonghuJeon/3da-libero-training-assets into the repository root to provide the DA3-Giant base checkpoint and LIBERO training assets.

Project page: https://cvlab-kaist.github.io/Geometric-Action-Model/ Paper: https://arxiv.org/abs/2606.17046 Code: https://github.com/JeonSeongHu/gam_test/tree/public/libero-da3giant