Faster-WAM

This repository contains the Faster-WAM policy checkpoints for LIBERO and RoboTwin 2.0. Faster-WAM uses a Wan2.2 video backbone and a single-layer action expert with Lite KV-Fusion and Retracted 1D RoPE.

The checkpoints use the format supported by the official Faster-WAM code, release commit a66d16e.

Checkpoints

Benchmark Policy checkpoint Training step Normalization statistics
LIBERO / LIBERO-Plus libero/faster_wam_libero.pt 21,700 libero/dataset_stats.json
RoboTwin 2.0 robotwin/faster_wam_robotwin.pt 58,710 robotwin/dataset_stats.json

The LIBERO checkpoint and statistics are also used for LIBERO-Plus evaluation. Each checkpoint is approximately 10.06 GB. Use the statistics from the same benchmark directory as the checkpoint.

Download

After installing the official code and its dependencies, run the following from the code repository root:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="TeamFasterwam/FasterWAM",
    local_dir="checkpoints/faster_wam",
)

To download one benchmark only, add allow_patterns=["libero/*", "README.md", "LICENSE"] or replace libero with robotwin.

Evaluation

Install and configure the corresponding simulator environment using the LIBERO / LIBERO-Plus guide or the RoboTwin guide. Run each command below from the code repository root, with the matching simulator environment activated.

LIBERO:

source scripts/env.sh libero
bash scripts/eval.sh libero model=fasterwam \
  ckpt="$PWD/checkpoints/faster_wam/libero/faster_wam_libero.pt" \
  EVALUATION.dataset_stats_path="$PWD/checkpoints/faster_wam/libero/dataset_stats.json"

For LIBERO-Plus, activate its separate simulator environment and replace libero with libero_plus in the two script arguments; keep the LIBERO checkpoint and statistics paths.

RoboTwin 2.0:

source scripts/env.sh robotwin
bash scripts/eval.sh robotwin model=fasterwam \
  ckpt="$PWD/checkpoints/faster_wam/robotwin/faster_wam_robotwin.pt" \
  EVALUATION.dataset_stats_path="$PWD/checkpoints/faster_wam/robotwin/dataset_stats.json"

The simulator assets, VAE, text encoder and tokenizer are installed or downloaded separately as described in the code repository. These are custom PyTorch policy checkpoints loaded by Faster-WAM, rather than a Transformers from_pretrained model. They contain policy weights and do not include optimizer state for resuming training.

Checkpoint format and verification

The release conversion preserved all tensor values while updating parameter names and checkpoint metadata. Each benchmark's conversion.json records the source and release SHA-256 hashes, training step and model options. Both models use readout=lite_kv_fusion, key_aggregation_order=norm_rope_then_fuse, position_interface=retracted_1d, and action_text_cross_attn=false. The camera layout is horizontal for LIBERO and robotwin for RoboTwin.

SHA256SUMS lists checksums of the files in this release. To verify a complete download:

cd checkpoints/faster_wam
sha256sum -c SHA256SUMS

License

Faster-WAM's original contributions are provided under the PolyForm Noncommercial License 1.0.0, matching the code release. Third-party components retain their own licenses. The accompanying third-party notices are copied from the code release; the source-code locations in that document refer to the GitHub repository.

Downloads last month

-

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

Model tree for TeamFasterwam/FasterWAM

Finetuned
(145)
this model