Official implementation of SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning.
SAGE plans with a frozen latent world model. A subgoal generator predicts intermediate targets, an action prior proposes goal-directed action sequences, and the world model evaluates and refines them before execution.
conda env create -f environment.yml
conda activate sage
pip install -e .Run from the repository root on Linux. LIBERO and RoboTwin use separate simulator environments.
Prepare the PushT Lance and OGBench Cube HDF5 datasets, then run:
PUSHT_DATASET=/path/to/pusht_expert_train.lance \
CUBE_DATASET=/path/to/cube_single_expert.h5 \
bash scripts/reproduce_main.shThis downloads the original SAGE component checkpoints, runs the PushT/Cube
component evaluations, and saves results under results/.
Fixed splits and query manifests are included in data/.
To evaluate selected methods:
export PUSHT_DATASET=/path/to/pusht_expert_train.lance
METHODS="base_cem sage" bash scripts/eval_pusht.sh| Experiments | Instructions |
|---|---|
| PushT / OGBench Cube, visual-only, DINO-WM | Protocols |
| LIBERO / RoboTwin | Download and reproduce / Detailed setup |
| Datasets and manifests | Data layout |
bash scripts/train_pusht.sh
bash scripts/train_cube.shThese recipes train the subgoal generator and action priors on a frozen world
model. Architectures and sampling settings are in configs/training.json;
additional recipes are documented in Protocols
and Native Suites.
RoboTwin A2B collection and end-to-end world-model training are described in
the A2B recipe.
@article{cheng2026sage,
title={SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning},
author={Cheng, Letian and Zhang, Qi and Wang, Qixun and Wang, Yisen},
journal={arXiv preprint arXiv:2607.17973},
year={2026}
}