StarWAM adapter model card

The machine-readable capability declaration is pinned in configs/models/starwam-capabilities-v0.1.json and validated against the public Draft 2020-12 capability schema in CI.

Model and adapter identity

Field Value
WAMProbe model ID starwam-wan22-5b-mot-libero
Upstream code shaohua-pan/StarWAM@f6c771fc3be0a9bc271ea4f1531d8ea35efb0ec7
Released model panshaohua/starwam@7d4bfe3ec76172ca17169fa959d21da099d386fe
Checkpoint SHA256 d24edea01579880327cfd9dc84d24adab82e420dca9652e614ad697bc8cc5378
Backbone Wan2.2 TI2V-5B revision 921dbaf3f1674a56f47e83fb80a34bac8a8f203e
Adapter version 0.1.0
Upstream code/model license Apache-2.0 as declared by upstream repositories

The WAMProbe adapter is an isolated integration target, not a reimplementation or an endorsement by the StarWAM authors. Weights remain outside Git.

Verified interface

Input:

  • two ordered 256×256 RGB cameras (agentview, wrist), rotated 180° and resized to 224×224 before horizontal concatenation;
  • eight proprioceptive values: EEF position, axis-angle orientation and two gripper positions;
  • the LIBERO natural-language task, encoded with the pinned UMT5-XXL text encoder;
  • an explicit non-negative diffusion seed.

Output:

  • 32 × 7 denormalized LIBERO delta-pose/gripper action chunk;
  • measured latency, device, dtype and peak allocated/reserved GPU memory;
  • full code/model/preprocessing/inference provenance and a prediction payload SHA256.

Declared capabilities

The released integration is verified for observation/task → action prediction. It does not expose an action-conditioned predict_future(candidate_action) endpoint, future RGB video, future state, uncertainty or likelihood through the WAMProbe adapter. Therefore:

  • StarWAM action chunks can be executed and scored in LIBERO;
  • NFE, seed and executed-prefix horizon can be evaluated;
  • candidate-action mask/shuffle, PSNR/SSIM/FVD and future-state ADE/FDE are skipped with structured reasons for this adapter;
  • internal video/world features are not relabeled as predicted videos.

Verified evaluation

The 2026-07-15 matrix used four LIBERO-CF-Mini task families, seeds 0/1/2 and NFE 1/4/8: 36 predictions and 36 action executions completed with zero runtime failures. A second inference pass returned 36 SHA-verified cache hits. Mean latency rose from 0.780 s at NFE 1 to 1.216 s at NFE 8 while peak allocation remained about 11.39 GiB.

Actions were executed from the same content-addressed snapshots with checkpoints after 8/16/32 controls. Sparse success and return were zero for every fixed context. Mean EEF displacement was 0.1716/0.1121/0.0000 m, showing that a longer open-loop action prefix did not monotonically improve control. This is a negative short-horizon diagnostic result, not a LIBERO policy success-rate estimate.

See the experiment report and LIBERO-CF-Mini benchmark card.

Limitations and risks

  • Four fixed init states cannot estimate general LIBERO success.
  • The checkpoint was evaluated without retraining or calibration.
  • Open-loop prefixes can leave the training distribution; WAMProbe does not assert robot safety or constraint satisfaction.
  • Seed/NFE sensitivity is descriptive with only three values per axis.
  • The current adapter cannot isolate whether internal future features caused an action.
  • Results depend on pinned CUDA/PyTorch/MuJoCo hardware and software; GPU nightly evidence should be treated as a reproducibility check, not numerical bit-equivalence across GPUs.