You can deploy this model and start earning money today!

SaltedLemon/
lejepa-control-pusht

Name: lejepa-control-pusht
Availability: InStock

apache-2.0

Views

No views yet

LeJEPA Control checkpoints

Trained controllers, planners, frozen LeWM world models, latent caches, and manifold-transfer experiments for lejepa_control. The controller and planner checkpoints are trained in this project; the LeWM world model is always frozen while those planners are trained.

Main PushT assets

checkpoints/controller/controller.pt: original phase-1 iterative controller.
checkpoints/ablations/: terminal-only, support-loss, and arrival+hold ablations.
checkpoints/planner_*: recursive and cross-attention planners.
checkpoints/density/density.pt: PushT behavior-density model.

The headline PushT checkpoint is checkpoints/ablations/ah_hold0.5/controller.pt (94% at receding horizon 1 on 50 held-out episodes, seed 42). Raw percentages are only comparable under the same checkpoint, harness, start/goal pairs, and receding-horizon setting; use the repository's paired McNemar/bootstrap tooling for comparisons.

Environment-native controllers

checkpoints/environment_controllers/<environment>/ contains a matched pair:

environment	files	notes
`tworooms`	`controller.pt`, `density.pt`	Native TwoRoom controller; 98% in the E1 50-episode run.
`reacher`	`controller.pt`, `density.pt`	Native DMC Reacher controller; 98% in the E1 50-episode run.
`humanoid`	`controller.pt`, `density.pt`	Experimental DMC Humanoid controller. The old 16% figure used a broken fallen-state eval set and must not be treated as walking performance.

Each controller checkpoint embeds its training arguments, action statistics, step, and validation profile.

Frozen base world models

checkpoints/base_world_models/<environment>/ contains configs and weights.

pusht, reacher, and tworooms mirror the official quentinll/lewm-* releases for reproducible, self-contained experiments. Their upstream model cards are included as UPSTREAM_README.md; those upstream artifacts are MIT licensed.
humanoid/weights_epoch_40.pt is the project-trained 84x84 Humanoid LeWM checkpoint. It is experimental and uses a different patch grid from the 224x224 released models.

Manifold-transfer checkpoints

checkpoints/manifold_transfer/e2_adapters/: post-hoc TwoRooms/Reacher to PushT latent maps, with and without the 10-to-10 action adapter.
checkpoints/manifold_transfer/e3_lite/: frozen-backbone projector runs. pusht_noB.pt and pusht_withB.pt are the two zero-shot transfer variants that reached 70% and 68% on TwoRoom with the unchanged PushT controller.
checkpoints/manifold_transfer/e3_full/: full encoder-training runs, grouped by condition and architecture. best.pt is validation-selected, final.pt is the last training step, and the main PushT/ResNet18 convergence snapshots use zero-padded step names.

The E3-lite result is partial alignment, not parity: the native TwoRoom stack reached 98%, while unrelated or negative-control stacks sat around the 24-38% floor. See the repository's manifold_transfer/results/SYNTHESIS.md for exact paired statistics, controls, and caveats.

Loading

These are native PyTorch checkpoints rather than a Transformers pipeline:

python

import torch

checkpoint = torch.load("controller.pt", map_location="cpu", weights_only=False)
print(checkpoint["args"])

Use the repository code to rebuild the matching architecture before loading state_dict. Treat PyTorch pickle checkpoints as executable input and only load files from sources you trust.

Licenses and attribution

Project code and project-trained artifacts follow the repository's Apache-2.0 licensing. Mirrored official LeWM PushT/Reacher/TwoRooms artifacts retain their upstream MIT license and attribution to the LeWorldModel authors. Datasets are not included in this model repository.

SaltedLemon/lejepa-control-pusht