e05 embodied-pretrained VGGT β€” step 260,000

Source run: beta_joint_7n_2ep_cosine_reader_w12_v1. Source checkpoint: step_260000.pt. Source SHA256: e35afc81970c8b8938695b866b9911c503520d5ccb0a2f2ed37a3efcfc69cf0f. Export SHA256: ce5698a438b3f246d1d1fe0bcc4d1b205876c0523066cf11f08f8434d98770c5.

This resource initializes the VGGT branch for real-robot training. RoboTwin training uses the public VGGT-Omega initialization instead.

Includes aggregator, language projector, camera head and dense head: 1,415 tensors, 1,341 FP32 and 74 BF16, all preserved byte-for-byte. No action expert, bridge, proprio encoder or optimizer state is included. File size: 4,597,332,135 bytes.

Load with model.vggt_backbone_pretrained_checkpoint_path; the e05 repository provides +model_init=embodied_vggt and E05_EMBODIED_VGGT_CHECKPOINT. The payload format is e05_embodied_vggt, version 1, with a complete vggt_backbone state dictionary. This is not a complete inference policy.

Retain upstream VGGT-Omega model/code terms, including noncommercial research terms, and the e05 repository's THIRD_PARTY_NOTICES.md. The code's MIT license does not relicense model weights. Real-robot task configurations and evaluation are separate from extraction of this resource.

Download and training initialization

Run from the e05 code repository:

source scripts/env.sh
hf download X-EraAI/e05-embodied-vggt --local-dir checkpoints/embodied-vggt
(cd checkpoints/embodied-vggt && sha256sum -c SHA256SUMS)
export E05_EMBODIED_VGGT_CHECKPOINT="$E05_ROOT/checkpoints/embodied-vggt/e05_embodied_vggt_step_260000.pt"

Append +model_init=embodied_vggt to your real-robot Hydra training command, with your task's data, normalization, and action/state dimensions. Do not pass this file as a full-model inference checkpoint or resume state. RoboTwin training retains public VGGT initialization.

Resources

Terms

See third-party notices. The code repository's MIT license does not relicense weights or data.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading