YunzeLiu/fastwam-trex62-baseline

FastWAM T-Rex 62-D Baseline — Step 27,000

This repository contains the selected late-training FastWAM baseline checkpoint for the pick_fruits split of wushr-lance/VLA2Vec. It replaces the previously published step-4,500 early-training checkpoint.

Status: checkpoint at step 27,000 of the configured 30,000-step schedule (90.0%). This is a consolidated model-weight release for inference and checkpoint-stage comparisons. It has not yet received a T-Rex real-robot rollout evaluation, so the offline metrics below must not be interpreted as a robot success rate.

Checkpoint

Field	Value
File	checkpoints/step_027000.pt
Size	12,042,403,349 bytes
SHA-256	9c79459497cfacd19e9e1037bb3d3293da2a6167e2667c61bf8c1c3130bba683
Run ID	trex62_baseline_16gpu_v3
Training step	27,000 / 30,000
Epoch	39
Global batch size	128
Precision	BF16

Only consolidated model weights are published. DeepSpeed optimizer, scheduler, sampler, and random-state files are not included.

Model and data contract

• Architecture: Wan2.2-TI2V-5B video DiT plus ActionDiT in FastWAM's mixture-of-tokens setup.
• Dataset: 100 pick_fruits episodes, 92,594 frames, three RGB cameras.
• Visual canvas: 448 x 384, composed from one head view and two wrist views.
• Action chunk: 32 steps.
• Action/proprioception contract: T-Rex 62-D bimanual action space.
• 62-D layout: left arm 9 + left hand 22 + right arm 9 + right hand 22.
• Arm rotation representation: the first two columns of the rotation matrix.
• Dataset normalization: exact q01/q99 min-max statistics to [-1, 1].

The source dataset revision used by this run is e7bbf566b881414beaee90507888cb47d5d17b7f.

Step-27,000 offline snapshot

Metric	Value
Training total loss	0.033321
Training video loss	0.032672
Training action loss	0.000649
Validation loss	0.289092
Validation action L1	0.027688
Validation action L2	0.002799

The repository includes artifacts/metrics_through_step_027000.jsonl, the run configuration captured from the training directory, and the exact normalization statistics. These are offline training/validation measurements, not real-robot rollout results.

Download and integrity check

huggingface-cli download YunzeLiu/fastwam-trex62-baseline \
  --local-dir fastwam_trex62_baseline
sha256sum fastwam_trex62_baseline/checkpoints/step_027000.pt

The expected digest is 9c79459497cfacd19e9e1037bb3d3293da2a6167e2667c61bf8c1c3130bba683.

Reproduction and loading

Use the configuration, normalization statistics, and metric history included in this repository. The matching implementation is available at yunzeliu/fastwam_icl. The checkpoint uses FastWAM's standard consolidated-weight loading path.

Limitations

• No T-Rex real-robot rollout evaluation is included in this release.
• The checkpoint contains model weights only, not resumable optimizer state.
• It is specialized to the fixed pick_fruits data, action, and camera contract above.

Licensing

The FastWAM integration code is MIT licensed. Base-model weights, third-party components, and the VLA2Vec dataset retain their respective licenses and terms.