feat(gym): allow configured sensors to be disabled

[yuecideng]

Description

This PR adds an opt-out path for sensor acquisition in Gym deployments.

• Add enable_sensor: false to keep an embodiment sensor declaration while skipping sensor creation and image acquisition.
• Add --disable_sensor / --disable-sensor launcher overrides.
• Omit sensor image allocation from the online-data rollout buffer when acquisition is disabled.
• Remove sensor-dependent event and observation functors before their managers are initialized, including nested SceneEntityCfg, all_sensors, sensor output paths, and camera-keyed descriptors.
• Add focused coverage using the existing CobotMagic tableware embodiment and manager-level tests.

Dependencies: None.

Validation

• pytest -q tests/gym/envs/test_sensor_acquisition.py tests/gym/utils/test_gym_utils.py — 132 passed
• black --check --diff --color ./
• python docs/scripts/check_api_docs.py
• python .agents/skills/project-dev-context/scripts/context.py check

The Python 3.10 environment could not collect pytest because its installed SciPy/Scikit-learn binaries have an ABI signature mismatch; the same tests pass under Python 3.11.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Adds optional sensor disable switch to environment configuration.

The PR appears safe to merge, though disabled sensor add-mode observations can still consume rollout-buffer memory.

Findings

1.  Disabled images still allocated ▶

Fix with agent prompt

### Issue 1
embodichain/lab/gym/utils/gym_utils.py:1515
When a disabled sensor has an `add`-mode observation with an image-shaped `extra.shape`, the buffer allocates that observation before this sensor check. Environment initialization removes the observation functor, so the allocated image buffer cannot be populated and can retain the memory cost this switch is meant to avoid. Please apply the disabled-sensor check to extra-observation allocation too.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

The PR adds a Gym configuration and launcher switch to skip configured sensor acquisition, filters sensor-dependent event and observation functors, and omits configured sensor entries from the online rollout buffer.

• Adds focused tests and updates environment-framework guidance.
• Image-shaped extra observations can still be allocated when their sensor functors are disabled.

_{Reviews (1) · Last reviewed commit: "feat(gym): disable configured sensor acq..."}

[greptile-apps]

Disabled images still allocated

When a disabled sensor has an add-mode observation with an image-shaped extra.shape, the buffer allocates that observation before this sensor check. Environment initialization removes the observation functor, so the allocated image buffer cannot be populated and can retain the memory cost this switch is meant to avoid. Please apply the disabled-sensor check to extra-observation allocation too.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/lab/gym/utils/gym_utils.py
Line: 1515

Comment:
**Disabled images still allocated**

When a disabled sensor has an `add`-mode observation with an image-shaped `extra.shape`, the buffer allocates that observation before this sensor check. Environment initialization removes the observation functor, so the allocated image buffer cannot be populated and can retain the memory cost this switch is meant to avoid. Please apply the disabled-sensor check to extra-observation allocation too.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.