[Task Clean-up][Manager][Camera] Dexterous Part 10/11: Add the Shadow camera manager counterpart

[hujc7]

Summary

• Adds the manager-based counterpart for the Shadow Hand camera reorientation task (Play variant, benchmark row, rendering-correctness row, CI-rendered golden images for both camera tasks).
• Carries the camera enablement deferred from Part 6 per review: the shared validation helper, the Warp cube-keypoint implementation with parity tests (the module-level compute_keypoints is deprecated in favor of feature_extractor.compute_cube_keypoints), and the camera observation terms in the task mdp package.
• The benchmark row gates on episode length and duration only: the camera reward scale plateaus near 300 against the copied 1000 gate, and camera threshold recalibration is a tracked follow-up. The consumer-less manager Benchmark registration was dropped per review.
• Validated by full camera manager training and rendering-correctness runs; split out of the lumped validation branch [DO-NOT-MERGE][Task Clean-up] Dexterous: lumped validation branch (series reference) #6324 (Part 10 of 11).

Dependencies

• Code: [Task Clean-up] Dexterous Part 3/9: Add success-rate metrics to the reorientation Direct tasks #6413 (Part 3), [Task Clean-up][Camera] Dexterous Part 6/11: Add renderer presets to the Direct camera task #6416 (Part 6), [Task Clean-up][Manager] Dexterous Part 5/9: Add the reorientation manager counterparts #6418 (Part 8), [Task Clean-up][Manager] Dexterous Part 9/11: Add Shadow state and OpenAI manager counterparts #6419 (Part 9 — Shadow manager base). CI stays red until they merge; the branch will be refreshed then.
• Runtime: [Task Clean-up][Newton] Dexterous Part 1/8: Fix cubric fallback and visualizer teardown #6411 (Part 1) for the Newton renderer column.

Review updates (2026-07-09)

• The camera manager reads fingertip names and the force-torque scale from the constants module instead of instantiating the Direct cfg; the cube-keypoint feature path writes into a pre-allocated buffer.

Review updates (2026-07-11)

• Synced with lump head d3cddae: the camera manager benchmark row gains an interim reward gate (150) that catches pipeline breakage below the truncated-run plateau (~174-290); success gates stay off until a full-budget manager run calibrates them. The camera observation terms and feature extractor pick up the Warp kernel conversion, and the camera env drops its redundant fingertip force-sensor update. The keypoint parity test file is dropped; the consolidated kernel behavior suite ships with Part 11 ([Task Clean-up][Manager] Dexterous Part 6/9: Add the handover and camera manager counterparts #6421).

Series review map

Full integrated diff + training/validation evidence: the lumped validation PR #6324 (DO-NOT-MERGE).

Part	PR
Docs: regenerate the environment overview table	#6410
Part 1/11: Newton runtime fixes (cloner rows, cubric fallback, viz teardown)	#6411
Part 2/11: OVPhysX runtime fixes	#6412
Part 3/11: success-rate metrics for the Direct reorientation tasks	#6413
Part 4/11: RSL-RL training for the handover Direct task	#6414
Part 5/11: success-rate support in the benchmark utilities	#6415
Part 6/11: renderer presets for the Direct camera task	#6416
Part 7/11: OVPhysX presets for the dexterous tasks	#6417
Part 8/11: Allegro manager counterpart	#6418
Part 9/11: Shadow + OpenAI manager counterparts	#6419
Part 10/11: Shadow camera manager counterpart (this PR)	#6420
Part 11/11: Shadow handover manager counterpart	#6421

Builds on: #6413, #6416. This PR's own commits relative to each parent branch:

• vs part 6413: hujc7/IsaacLab@jichuanh/task-cleanup-dex-part03...jichuanh/task-cleanup-dex-part10
• vs part 6416: hujc7/IsaacLab@jichuanh/task-cleanup-dex-part06...jichuanh/task-cleanup-dex-part10

[greptile-apps]

Greptile Summary

This PR adds the manager-based counterpart for the Shadow Hand camera reorientation task, completing parity with the existing Direct variant. It introduces ShadowHandCameraManagerEnvCfg (plus Play and Benchmark variants), three new gym registrations, a benchmark row, render-correctness test registration, and golden images for both the Direct and manager camera tasks.

• New shadow_hand_camera_manager_env_cfg.py: Wires a _ShadowHandCameraManagerSceneCfg (camera + joint-wrench sensors) together with asymmetric actor/critic observations (CameraPolicyCfg / CameraCriticCfg) that mirror the Direct task's 191-dim / 214-dim observation spaces, sharing the same FeatureExtractorCfg instance via __post_init__.
• __init__.py additions: Registers Isaac-Reorient-Cube-Shadow-Camera, -Play, and -Benchmark for ManagerBasedRLEnv, and also back-fills the missing rsl_rl_cfg_entry_point on the existing Isaac-Reorient-Cube-Shadow-OpenAI-LSTM-Direct entry.
• Test/benchmark additions: The training env is registered in the render-correctness suite with the same flaky(max_runs=3, min_passes=1) marks as the Direct camera task, and benchmark thresholds match the Direct task's values.

Confidence Score: 4/5

Safe to merge once the dependency chain (Parts 3, 6, 8, 9) lands; the new config code is structurally consistent with the Direct camera task it mirrors.

The new manager env cfg correctly reuses Direct task dynamics and wires asymmetric actor/critic observations. Two style-level concerns exist: validate_config hardcodes the three scene preset attribute names rather than discovering them dynamically, and a full ShadowHandCameraEnvCfg() is instantiated at module import time just to read two constants. Neither represents a functional defect in the changed code; the PR was validated end-to-end with full training runs.

shadow_hand_camera_manager_env_cfg.py — the validate_config method and the module-level config instantiation are worth a second look.

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/shadow_hand_camera_manager_env_cfg.py	New manager-based camera env config reusing Direct task's dynamics and observations; structurally sound with a minor coupling concern in validate_config and module-level instantiation.
source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/init.py	Adds gym registrations for three camera manager variants (training, play, benchmark) and fills in missing rsl_rl runner for the existing LSTM-Direct env; consistent with existing patterns.
source/isaaclab_tasks/test/benchmarking/configs.yaml	Adds benchmark row for Isaac-Reorient-Cube-Shadow-Camera with thresholds identical to the Direct camera task; consistent with the series pattern.
source/isaaclab_tasks/test/core/test_rendering_registered_tasks.py	Registers Isaac-Reorient-Cube-Shadow-Camera in the render-correctness suite with the same flaky-retry marks as the Direct camera task; no logic changes to the test harness.

Class Diagram

%%{init: {'theme': 'neutral'}}%%
classDiagram
    class ShadowHandManagerEnvCfg {
        +scene: _ShadowHandManagerSceneCfg
        +observations: FullStateObsCfg
    }
    class ShadowHandCameraManagerEnvCfg {
        +scene: ShadowHandCameraManagerSceneCfg
        +observations: CameraObservationsCfg
        +feature_extractor: FeatureExtractorCfg
        +validate_config()
        +__post_init__()
    }
    class ShadowHandCameraManagerPlayEnvCfg {
        +scene: ShadowHandCameraManagerPlaySceneCfg
        +feature_extractor: FeatureExtractorCfg(train=False, load_checkpoint=True)
    }
    class ShadowHandCameraManagerBenchmarkEnvCfg {
        +feature_extractor: FeatureExtractorCfg(enabled=False)
    }
    class CameraObservationsCfg {
        +policy: CameraPolicyCfg
        +critic: CameraCriticCfg
    }
    class CameraPolicyCfg {
        +camera_features: ObsTerm(ShadowHandCameraFeatures)
        +goal_keypoints: ObsTerm
        +last_action: ObsTerm
        --object_pos/quat/vel set to None
    }
    class CameraCriticCfg {
        +fingertip_wrench: ObsTerm
        +last_action: ObsTerm
        +camera_features: ObsTerm(cached)
    }
    class _ShadowHandCameraManagerSceneCfg {
        +ground = None
        +tiled_camera: ShadowHandTiledCameraCfg
        +joint_wrench: JointWrenchSensorCfg
    }

    ShadowHandManagerEnvCfg <|-- ShadowHandCameraManagerEnvCfg
    ShadowHandCameraManagerEnvCfg <|-- ShadowHandCameraManagerPlayEnvCfg
    ShadowHandCameraManagerEnvCfg <|-- ShadowHandCameraManagerBenchmarkEnvCfg
    ShadowHandCameraManagerEnvCfg *-- CameraObservationsCfg
    CameraObservationsCfg *-- CameraPolicyCfg
    CameraObservationsCfg *-- CameraCriticCfg
    ShadowHandCameraManagerEnvCfg *-- _ShadowHandCameraManagerSceneCfg

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"}}}%%
classDiagram
    class ShadowHandManagerEnvCfg {
        +scene: _ShadowHandManagerSceneCfg
        +observations: FullStateObsCfg
    }
    class ShadowHandCameraManagerEnvCfg {
        +scene: ShadowHandCameraManagerSceneCfg
        +observations: CameraObservationsCfg
        +feature_extractor: FeatureExtractorCfg
        +validate_config()
        +__post_init__()
    }
    class ShadowHandCameraManagerPlayEnvCfg {
        +scene: ShadowHandCameraManagerPlaySceneCfg
        +feature_extractor: FeatureExtractorCfg(train=False, load_checkpoint=True)
    }
    class ShadowHandCameraManagerBenchmarkEnvCfg {
        +feature_extractor: FeatureExtractorCfg(enabled=False)
    }
    class CameraObservationsCfg {
        +policy: CameraPolicyCfg
        +critic: CameraCriticCfg
    }
    class CameraPolicyCfg {
        +camera_features: ObsTerm(ShadowHandCameraFeatures)
        +goal_keypoints: ObsTerm
        +last_action: ObsTerm
        --object_pos/quat/vel set to None
    }
    class CameraCriticCfg {
        +fingertip_wrench: ObsTerm
        +last_action: ObsTerm
        +camera_features: ObsTerm(cached)
    }
    class _ShadowHandCameraManagerSceneCfg {
        +ground = None
        +tiled_camera: ShadowHandTiledCameraCfg
        +joint_wrench: JointWrenchSensorCfg
    }

    ShadowHandManagerEnvCfg <|-- ShadowHandCameraManagerEnvCfg
    ShadowHandCameraManagerEnvCfg <|-- ShadowHandCameraManagerPlayEnvCfg
    ShadowHandCameraManagerEnvCfg <|-- ShadowHandCameraManagerBenchmarkEnvCfg
    ShadowHandCameraManagerEnvCfg *-- CameraObservationsCfg
    CameraObservationsCfg *-- CameraPolicyCfg
    CameraObservationsCfg *-- CameraCriticCfg
    ShadowHandCameraManagerEnvCfg *-- _ShadowHandCameraManagerSceneCfg

Loading

_{Reviews (1): Last reviewed commit: "Add the Shadow camera manager counterpar..." | Re-trigger Greptile}

[greptile-apps]

Module-level config instantiation to read two constants

ShadowHandCameraEnvCfg() is instantiated at import time solely to read fingertip_body_names and force_torque_obs_scale. Since both values are defined directly on ShadowHandEnvCfg (the parent class) as class-level fields, they can be retrieved without creating a full env config object. The current approach works but adds unnecessary initialization cost on every import of this module, and any future side effect added to ShadowHandCameraEnvCfg.__post_init__ would silently fire at import time.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

validate_config hardcodes preset attribute names

The three backend names physx, newton_mjwarp, and ovphysx are accessed by string-literal attribute lookup. All concrete subclasses in this PR define those three attributes, so this is safe today. However, any future subclass that overrides scene with a PresetCfg that omits or renames one of these backends will hit an AttributeError from PresetCfg.__getattr__ (which only handles known legacy aliases). Iterating via dataclasses.fields() and filtering out default and underscore-prefixed entries would make the validation self-describing and forward-compatible.

[hujc7]

Closing as part of the series restructure: the Shadow camera manager counterpart now lives in #6421 (Part 6/8). The full series layout is tracked in the Review Map at the top of #6324.