LeWM PushT Checkpoint Collection

This repository contains reproduced LeWM checkpoints for PushT across frame skips and predictor capacities.

These are reproduction checkpoints, not official checkpoints from the LeWM authors.

Contents

Checkpoints are organized by predictor family and frame skip:

vanilla_vits/fs5/
vanilla_vits/fs10/
vanilla_vits/fs20/
vanilla_vitb/fs5/
vanilla_vitb/fs10/
vanilla_vitb/fs20/

Each subdirectory contains:

file purpose
lewm_weights.ckpt Lightning training checkpoint
lewm_epoch_*_object.ckpt Object checkpoint used by native CEM eval
config.yaml Training configuration snapshot
hydra/.hydra/* Hydra-resolved training config
eval/native_results/*.json Native PushT CEM evaluation outputs

Training Setup

Common settings:

item value
environment PushT
source repo third_party/le-wm at commit ca231f9f
encoder ViT-Tiny, trained from scratch
epoch target anchor epoch 10
effective batch 128
precision FP32
hardware RLL02, RTX 2080 Ti DDP

Predictor settings:

family predictor trainable params batch
vanilla_vits ViT-S-style, hidden dim 384, depth 12, heads 6 about 42M total 32 x DDP4
vanilla_vitb ViT-B-style, hidden dim 768, depth 12, heads 12 about 140.6M total 32 x DDP4

Frame-skip step budgets:

fs max steps
5 154,822
10 125,627
20 68,997

Native PushT Planning Evaluation

The reported success rates use the LeWM paper-style native PushT planning protocol: sample start and goal states from the offline pusht_expert_train trajectories, then run CEM-MPC in the PushT simulator.

Native planning config:

item value
eval episodes 50 per eval seed
eval seeds 42, 123, 7
goal offset 25
horizon 5
action block frame skip
receding horizon 5
eval budget 50 for fs=5 and fs=10, 100 for fs=20
CEM samples 300
CEM iterations 30
CEM top-k 30
success criterion PushT pose tolerance, position < 20 and angle < pi/9

Paper-train results:

predictor fs seed 42 seed 123 seed 7 mean ± std
ViT-S 5 96.0% 92.0% 96.0% 94.67 ± 2.31%
ViT-S 10 72.0% 76.0% 64.0% 70.67 ± 6.11%
ViT-S 20 50.0% 48.0% 54.0% 50.67 ± 3.06%
ViT-B 5 86.0% 90.0% 88.0% 88.00 ± 2.00%
ViT-B 10 70.0% 78.0% 74.0% 74.00 ± 4.00%
ViT-B 20 48.0% 38.0% 40.0% 42.00 ± 5.29%

The ViT-S fs=5 reproduction matches the LeWM paper's reported PushT result of 96.0 ± 2.83% within the reported band.

Held-out reference, seed 42 on pusht_common_eval:

predictor fs success rate
ViT-S 5 80.0%
ViT-S 10 62.0%
ViT-S 20 40.0%
ViT-B 5 80.0%
ViT-B 10 70.0%
ViT-B 20 42.0%

The held-out numbers are stricter references and are not the paper's native evaluation protocol.

Loading

Download the repo and point AutoCostModel to the desired subfolder:

from pathlib import Path
from huggingface_hub import snapshot_download
from stable_worldmodel.policy import AutoCostModel

repo_dir = Path(snapshot_download("MasonJK99/lewm-pusht"))
model = AutoCostModel(repo_dir / "vanilla_vits" / "fs10")

For Lightning checkpoint inspection:

from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(
    "MasonJK99/lewm-pusht",
    "vanilla_vitb/fs10/lewm_weights.ckpt",
)
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)

Notes

  • The native planner requires the LeWM/stable-worldmodel environment used for training and evaluation.
  • Evaluation JSONs are included so the reported success rates can be audited.
  • Dataset files are not included in this repository.
  • The LeWM source code is MIT licensed; dataset usage is subject to the PushT / stable-worldmodel dataset terms.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support