boeyyyy/EgoWAM-checkpoints

EgoWAM checkpoints

Checkpoints for EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data (CoRL 2026).

Project Page  |  Paper  |  arXiv  |  Code

Contents

Path	What it is
rae_stage1/ep-0000002.pt	RAE stage-1 decoder, the DINO variant's world-model target
policies/fold_clothes_dinov2_c_widetrunk_1799.ckpt	fold-clothes policy, DINOv2 stem, epoch 1799
robotwin_pick/*_1999.ckpt	eight RoboTwin pick policies, epoch 1999

The RoboTwin set covers four world-model targets × two regimes — hpt (BC), vae (Pixel), rae (DINO), 3dflow (3D Flow), each single (one embodiment) and cross (cross-embodiment). Held-out closed-loop success on the paper's fixed 100-seed list:

	BC s	BC c	Pixel s	Pixel c	DINO s	DINO c	3D Flow s	3D Flow c
success	2%	9%	6%	17%	3%	28%	1%	19%

Loading

These are training checkpoints, so they pickle references to modules that must be importable — including the pre-rename package name egomimic, which the repository aliases for exactly this reason:

import sys
sys.path.insert(0, "external/cosmos-policy")   # Pixel/DINO checkpoints reference it
sys.path.insert(0, "external/RAE/src")         # DINO checkpoints reference RAE's stage1
from egowam.compat import install_legacy_module_alias
install_legacy_module_alias()

import torch
ck = torch.load("robotwin_pick/rae_cross_1999.ckpt", map_location="cpu", weights_only=False)

Without the alias you get ModuleNotFoundError: No module named 'egomimic'.

To evaluate a policy rather than inspect it, use the repository's own entrypoint, which rebuilds the model from the checkpoint's embedded config:

python egowam/trainHydra.py --config-name=train_zarr_wm.yaml \
  +validate=true ckpt_path="<file>" data=<task>/<data_config>

Notes

• Raw training checkpoints. They carry optimizer state as well as weights, so they are larger than inference-only exports but can resume training.
• The RAE checkpoint holds model and ema. ema is normally the better of the two. Of its 502M parameters, 87M are the frozen facebook/dinov2-base encoder; the remaining 415M are the trained decoder.
• Third-party weights. The RAE decoder builds on RAE (MIT) and facebook/dinov2-base (Apache-2.0). The Pixel variant additionally needs NVIDIA's gated Cosmos tokenizer, which is not redistributed here — request access on its model page.

Citation

@article{li2026egowam,
  title   = {EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data},
  author  = {Li, Baoyu and Yin, Xinchen and Lin, Mengying and Zhang, Yixin and Xu, Danfei},
  journal = {arXiv preprint arXiv:2607.08436},
  year    = {2026}
}