FastWAM Optional-IDM β€” Songling, step 45,000

Intermediate model checkpoint at step 45,000. The recorded training target was 63,292 updates; this package is not the final one-epoch checkpoint. The weights were exported from a 32-GPU training state and contain mot, proprio_encoder, step, and torch_dtype. They do not include optimizer, scheduler, or RNG state for exact training resumption.

The local weight file has been loaded successfully on CPU, and its embedded step is 45,000. No robot rollout or held-out success rate is claimed.

Contents

  • weights/step_000045000.pt: BF16 model weights (12,041,707,939 bytes).
  • config.yaml: recorded training configuration.
  • dataset_stats.json: matching normalization statistics.
  • inference_contract.json: camera, joint and gripper conventions.
  • STAGE_COMPLETE.json: checkpoint step, dimensions and file sizes.
  • songling_stage_policy.py: loader for this intermediate checkpoint.
  • code/fastwam_source.zip: matching FastWAM source, including Optional-IDM.

The upload package removes the extra digest checks from the intermediate loader. File-size, step, architecture and strict parameter-loading checks remain. Model weights and normalization values are unchanged. Offline W&B records, training logs, optimizer shards and datasets are not included.

Loading

Extract the source archive from this repository root:

python -m zipfile -e code/fastwam_source.zip code

Install the dependencies of that FastWAM checkout, then load the complete repository while preserving its relative paths:

from songling_stage_policy import load_policy
policy = load_policy(
    "/absolute/path/to/downloaded/repository",
    device="cuda",
    action_infer_mode="idm",
    model_cache="/absolute/path/to/prepared/models",
)

The official Wan VAE, T5 encoder and tokenizer assets are required separately. The recorded configuration contains original training paths; the supplied loader selects the included trained weights and local normalization statistics.

Inputs are native uint8 RGB images named top_head, hand_left, hand_right, a 14-D robot state, and a task prompt. Outputs have shape (32, 14) and use absolute joint radians and native gripper metres. The adapter converts grippers using the recorded 0.08 m reference, which is a control-coordinate convention, not new hardware calibration. Images are resized to 224 square and concatenated horizontally into 224 x 672. Consult inference_contract.json before deployment.

Downloads last month
13
Video Preview
loading

Model tree for dddddsds31341/FastWAM-Songling-Optional-IDM-Step-45000

Finetuned
(142)
this model