Weights for the UniReps 2026 workshop extended abstract "Predicted Future Latents as Subgoals: Hierarchical Policies Outperform Flat Policies Beyond the Training Goal Distance" (Takahiro Morimoto, Miraiyochi, inc.).
Code: https://github.com/miraiyochi-ai/hierarchical-latent-subgoals
The hierarchical policy has two flow-matching levels and runs without test-time search. Every 16 environment steps the upper level generates the latent, in a frozen visual encoder's space, of the observation expected 16 steps ahead. The lower level takes that latent as its subgoal and generates the next 8 actions. Each hierarchical checkpoint comes with a flat policy trained on the same data with the number of trainable parameters matched within 0.1%.
| Path | Contents |
|---|---|
pusht_main_reproduction/seed_42/{hierarchical,flat_direct_goal}_ema.pt |
PushT, training goal distances d ∈ {25, 50, 75}, 30k updates, final EMA weights. Seed-42 rerun of the main setting (see note below). |
pusht_train150/seed_42/{hierarchical,flat_direct_goal}_ema.pt |
PushT range extension, training d ∈ {25, 50, 75, 100, 150}, seed 42. These are the weights reported in the paper. |
pusht_main_reproduction/seed_42/logs/, pusht_train150/seed_42/logs/ |
Records of these two runs, one folder per policy: training losses (training_losses.jsonl), effective configs, checkpoint summaries, evaluation logs and summaries for every subgoal condition and goal distance, and per-episode rollouts (evaluation_shards/ for d ≤ 75, h3x/ for d = 100 and 150). Paths inside the files refer to the training machines. |
pusht_data/ |
Fixed PushT inputs: train/eval split, in-range and beyond-range evaluation tasks, normalizers (--data-root of the code). |
libero_long/{hierarchical,flat}/ema_030000.pt |
LIBERO-Long (libero_10), frozen V-JEPA 2 ViT-L + T5-base, 30k updates, seed 42. These are the weights reported in the paper. |
libero_long/model_configs.json, language.json |
Model widths (parameter counts 74,846,727 vs. 74,854,568) and the frozen T5 instruction embeddings. |
Every folder has a SHA256SUMS file.
The paper's main PushT result averages three training seeds (42, 43, 44). Those six checkpoints were stored on pods without persistent storage and were not kept. The folder pusht_main_reproduction instead contains a later rerun of seed 42 with the same code, inputs and settings. Its results go the same way as the paper's (seed 42, hierarchical minus flat: +24 points at d = 100 and +26 points at d = 150, against +14 / +14 in the original seed-42 run). The values are not identical, so these weights reproduce the paper's per-seed numbers only approximately.
pusht_hi_lewm_epoch15_object.ckpt) is distributed with Hi-LeWM on Zenodo (DOI 10.5281/zenodo.21353240).facebook/vjepa2-vitl-fpc64-256) and T5-base (google-t5/t5-base) are on the Hugging Face Hub.@inproceedings{morimoto2026futurelatents,
title = {Predicted Future Latents as Subgoals: Hierarchical Policies Outperform Flat Policies Beyond the Training Goal Distance},
author = {Morimoto, Takahiro},
booktitle = {UniReps 2026 Workshop (Extended Abstract Track)},
year = {2026}
}
MIT.
libero_long/flat/ema_030000.pt 300 MBlibero_long/hierarchical/ema_030000.pt 299 MBpusht_main_reproduction/seed_42/hierarchical_ema.pt 271 MBpusht_train150/seed_42/hierarchical_ema.pt 271 MBpusht_main_reproduction/seed_42/flat_direct_goal_ema.pt 271 MBpusht_train150/seed_42/flat_direct_goal_ema.pt 271 MBpusht_data/s2_normalizer.npz 5.5 kBpusht_data/normalizers.npz 2.2 kB