[Task Clean-up][Camera] Dexterous Part 6/11: Add renderer presets to the Direct camera task

[hujc7]

Summary

• Adds an RGB-depth camera preset so the Shadow Hand camera environment can train with the Newton Warp renderer in addition to RTX (the only preset that is both Warp-renderable and CNN-trainable), with preset acceptance tests.
• Deprecates the Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct registration in favor of the env.feature_extractor.enabled=false override (no consumer exists; no other task family registers benchmark variants).
• Scope note: the camera manager enablement (shared validation helper, keypoint relocation, manager observation terms) lives in Part 10 per review.
• Validated by full camera training runs; split out of the lumped validation branch [DO-NOT-MERGE][Task Clean-up] Dexterous: lumped validation branch (split into Parts 1-11) #6324 (Part 6 of 11).

Dependencies

• None to merge. Runtime: the Newton-renderer validation column relies on [Task Clean-up][Newton] Dexterous Part 1/11: Fix cloner rows, cubric fallback, and visualizer teardown #6411 (Part 1).

[greptile-apps]

Greptile Summary

This PR adds an rgb_depth camera preset for the Shadow Hand camera environment, enabling training with the Newton Warp renderer using RGB + depth (4 channels). It also moves the cube-keypoint computation into a shared compute_cube_keypoints TorchScript function in feature_extractor.py and deprecates the old module-level compute_keypoints in shadow_hand_camera_env.py.

• New preset: ShadowHandTiledCameraCfg.rgb_depth sets data_types=[\"rgb\", \"depth\"]; warp-validation tests updated to include it.
• Refactor: Validation logic extracted to validate_shadow_hand_camera_settings; compute_keypoints becomes a warnings.warn shim delegating to the new function.
• New manager term infrastructure: ShadowHandCameraFeatures, shadow_hand_camera_cached_features, and shadow_hand_goal_keypoints are added to feature_extractor.py for future Manager-based observation use, but are not yet wired into any environment config in this PR.

Confidence Score: 4/5

The direct-env path (ShadowHandCameraEnv) is clean, and the preset/validation refactor is safe. The new ShadowHandCameraFeatures manager term has an AttributeError waiting to trigger when it is wired in.

The new ShadowHandCameraFeatures.call unconditionally calls .torch on physics-state tensors (root_pos_w, root_quat_w), which are plain torch.Tensor objects in Isaac Lab. torch.Tensor has no .torch attribute, so this will raise AttributeError the first time the class is exercised. Camera output is guarded correctly with isinstance, making the inconsistency clear. The class is not yet wired into any env config in this PR, so CI will not catch it now, but it will fail on first real use.

feature_extractor.py — specifically the ShadowHandCameraFeatures.call method at the object-state data access lines.

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/feature_extractor.py	Adds compute_cube_keypoints (TorchScript), ShadowHandCameraFeatures manager term, and two helper functions. The new manager term has a latent AttributeError on physics-state .torch access.
source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/shadow_hand_camera_env.py	Swaps direct compute_keypoints calls to compute_cube_keypoints; demotes old function to a deprecated shim with warnings.warn. Straightforward cleanup.
source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/shadow_hand_camera_env_cfg.py	Extracts validation logic into validate_shadow_hand_camera_settings and adds the rgb_depth preset. Behavior-equivalent refactor with a clean new preset.
source/isaaclab_tasks/test/core/test_shadow_hand_camera_presets.py	Adds rgb_depth to the valid-preset list and warp-valid-preset list; tests correctly cover the new preset.
source/isaaclab_tasks/changelog.d/task-cleanup-dex-part06.minor.rst	Changelog entry correctly documents the new rgb_depth preset and deprecation of compute_keypoints.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[ShadowHandCameraEnvCfg.validate_config] --> B[validate_shadow_hand_camera_settings]
    B --> C{tiled_camera is PresetCfg?}
    C -- Yes --> D[resolve to .default]
    D --> C
    C -- No --> E{renderer_cfg is PresetCfg?}
    E -- Yes --> F[resolve to .default]
    F --> E
    E -- No --> G{renderer_type == newton_warp?}
    G -- Yes --> H{unsupported data_types?}
    H -- Yes --> I[raise ValueError: unsupported types]
    H -- No --> J{depth-only + CNN enabled?}
    G -- No --> J
    J -- Yes --> K[raise ValueError: depth-only]
    J -- No --> L[valid config]

    subgraph New Presets
        M[ShadowHandTiledCameraCfg]
        M --> N[rgb: RGB only 3ch]
        M --> O[rgb_depth: RGB+depth 4ch NEW]
        M --> P[depth: depth-only benchmark]
        M --> Q[albedo/shading variants]
        M --> R[default/full: RGB+depth+seg 7ch]
    end

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[ShadowHandCameraEnvCfg.validate_config] --> B[validate_shadow_hand_camera_settings]
    B --> C{tiled_camera is PresetCfg?}
    C -- Yes --> D[resolve to .default]
    D --> C
    C -- No --> E{renderer_cfg is PresetCfg?}
    E -- Yes --> F[resolve to .default]
    F --> E
    E -- No --> G{renderer_type == newton_warp?}
    G -- Yes --> H{unsupported data_types?}
    H -- Yes --> I[raise ValueError: unsupported types]
    H -- No --> J{depth-only + CNN enabled?}
    G -- No --> J
    J -- Yes --> K[raise ValueError: depth-only]
    J -- No --> L[valid config]

    subgraph New Presets
        M[ShadowHandTiledCameraCfg]
        M --> N[rgb: RGB only 3ch]
        M --> O[rgb_depth: RGB+depth 4ch NEW]
        M --> P[depth: depth-only benchmark]
        M --> Q[albedo/shading variants]
        M --> R[default/full: RGB+depth+seg 7ch]
    end

Loading

_{Reviews (1): Last reviewed commit: "Add renderer presets to the Direct camer..." | Re-trigger Greptile}

[greptile-apps]

Unconditional .torch on physics-state tensors will raise AttributeError

object_asset.data.root_pos_w and root_quat_w return plain torch.Tensor objects — they come from PhysX, not from the Warp renderer. PyTorch tensors do not have a .torch attribute, so these two lines will raise AttributeError: 'Tensor' object has no attribute 'torch' the first time this class is exercised.

The camera output (a few lines below) handles this correctly with an isinstance guard, but the object-state path does not. The same defensive pattern is needed here.