fastwam-grasp-bag-strap-dr-v1

FastWAM grasp-bag-strap DR v1

FastWAM checkpoint fine-tuned without image augmentation on the deformable-bench/grasp_bag_strap_dr_v1 dataset.

Training setup

• Initialization: HWM FastWAM mid-training checkpoint at step 75,000
• Dataset revision: 239700e0cc56136cdc03a1d3625c1064e6fd48b9
• Dataset size: 200 episodes and 83,566 frames
• Cameras: static, left hand, and right hand
• Observation frames: 33
• Action/state dimension: 14
• Image preprocessing: resize to 240 x 320, without image augmentation
• Action/state normalization: min/max
• Batch size: 16 per GPU on 8 H20 GPUs
• DataLoader workers: 6
• Optimizer learning rate: 2e-5 with cosine decay
• Precision: bf16
• Training length: 3,265 optimizer steps

Training loss

The run completed normally at step 3,265. Mean logged losses over selected windows were:

The action loss converged strongly, while the late video and total losses remained stable with normal batch-to-batch variation.

Files

• checkpoints/step_003265.pt: final model weights
• config.yaml: resolved training configuration
• dataset_stats.json: normalization statistics used by training
• train.log: complete trainer log

Optimizer and distributed trainer state are intentionally excluded.

Intended use

This checkpoint is intended for research and evaluation with the matching FastWAM codebase and the grasp_bag_strap_dr_v1 observation/action schema. It has not been validated for safety-critical or out-of-distribution deployment.