LIBERO-Dynamic paper checkpoints
This repository contains the nine checkpoint versions used in the LIBERO-Dynamic dynamic-data training study, plus the native single-frame FastWAM-Joint checkpoint used for the paper's cross-model baseline comparison.
Contents
| Family | Variants |
|---|---|
| FastWAM-Joint, single frame | native baseline (step_034720.pt); static base; dynamic-only FT3000; dynamic+static FT3000 |
| FastWAM-Joint, history frame | dense-history static base; dynamic-only FT3000; dynamic+static FT3000 |
| ฯ0.5 | official static; dynamic-only FT1500; dynamic+static FT3000 |
The native baseline is stored at fastwam_joint/original_single_frame/step_034720.pt. It is distinct from fastwam_joint/single_frame/static_base/, the 60K-step static base used in the dynamic-training experiments. Its paired normalization file and provenance record are in the same directory.
The FastWAM files are PyTorch checkpoints and include matching dataset_stats.json files where they were part of the run. The ฯ0.5 entries contain inference parameters, checkpoint metadata, and normalization assets. Training optimizer state is intentionally omitted: this is an inference release, not a resume-training snapshot.
The original FastWAM-Joint checkpoint is released with its provenance retained and remains subject to the upstream FastWAM terms. Please cite LIBERO-Dynamic and the upstream FastWAM work when using it.
The history-frame reproduction package is also included under history_frame_fastwam/. It contains the history-frame model and loader snapshot, exact task/data/model configurations, and the aligned websocket adapter. The adapter consumes five observations and the preceding 16 executed actions as the clean prefix.
The two 3,000-step dynamic-only Joint checkpoints requested for reproduction are fastwam_joint/single_frame/dynamic_ft3000/step_003000.pt and fastwam_joint/history_frame/dynamic_ft3000/step_003000.pt; each directory contains its matching normalization statistics.
The corresponding strict non-benchmark training data is available at rsj2003/libero-dynamic-strict-nonbenchmark-1800. The complete code and evaluation protocol are in the LIBERO-Dynamic GitHub repository.