Add FastWAM RoboDojo evaluation adapter

[q6y6y6]

What changed

• add a lightx2v_ros/src/robodojo ament Python package with a reusable RoboDojo/XPolicyLab observation adapter for FastWAMPolicy
• map RoboDojo head and wrist RGB observations to the released RoboTwin FastWAM camera layout
• align evaluation defaults with the released baseline: 32-action chunks, 24 executed actions per plan, 10 denoising steps, and z-score normalization
• allow T5, tokenizer, and VAE assets to be provided through explicit absolute paths
• add optional sequential video/action expert offloading and transient CUDA cache cleanup for evaluation alongside Isaac Sim on a 24 GiB GPU
• avoid eagerly importing all optional pipeline dependencies when a model-specific runner is imported
• set the RoboTwin target video length used by the evaluated checkpoint

Why

This makes the native LightX2V FastWAM runtime directly reusable from the RoboDojo/XPolicyLab evaluation server, without falling back to the upstream FastWAM runtime. It also preserves the released action horizon/replanning behavior and supports the split asset layout used by the evaluation environment.

Evaluation status

The integration has completed end-to-end RoboDojo evaluation. Its current success rate is close to the upstream baseline: layout 0 succeeds, while layout 1 still fails. We suspect this layout-dependent behavior comes from the released official checkpoint; the current evidence is not sufficient to attribute the layout 1 result to the LightX2V integration.

Validation

• python -m compileall for the changed FastWAM modules and new adapter package
• python -m json.tool configs/fastwam/robotwin_i2va.json
• python setup.py check for the new ament Python package
• Python AST parse for the package sources
• git diff --check

pre-commit was not available in the local environment, so repository hooks were not run.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.