[mataylor-nvidia]

Description

Rename modifiers to tensor_processors and unify tensor and structured-data processing under TensorProcessor / TensorProcessorCfg. TensorProcessorData accepts Torch tensors, ProxyArray, and recursively nested lists, tuples, and string-keyed dictionaries. Each implementation defines its input and output structures; built-in filters retain their single-tensor behavior.

Custom processors can accept input buffers in their constructor and call super().__init__(cfg). Existing (cfg, data_dim, device) construction remains supported, and reset() and close() default to no-ops. Configurations use func for functions, classes, or lazy import strings; inputs and outputs provide optional named-key metadata. Managers resolve nested scene entities and manager terms in processor parameters.

Observation terms remain tensor-valued and shape-preserving. Groups retain their concatenated-tensor or dictionary-of-tensors output. Processing order, noise, clipping, scaling, delay, and history retain their existing behavior.

Legacy isaaclab.utils.modifiers and isaaclab_experimental.utils.modifiers imports remain supported with an import-time DeprecationWarning, including old class names, submodule paths, wildcard imports, and serialized callable paths. ObservationTermCfg.modifiers also remains supported: both observation managers emit a targeted FutureWarning and forward it to tensor_processors during initialization. Constructor arguments and later assignments work; setting both fields raises ValueError, including empty lists. Migrate imports and configurations to tensor_processors before the deprecated names are removed in a future release.

Camera attachment and Physically Plausible Image Signal Processing (PPISP) integration remain separate follow-up work. This PR establishes their shared processor interface and replaces the lower portion of #8571.

Validation

• uv run python -m pytest source/isaaclab/test/managers/test_observation_manager.py -q — 42 passed, including deprecated constructor/assignment use, serialized configurations, Warp manager integration, caller-directed warnings, and conflicting-field rejection.
• uv run --no-sync python -m pytest source/isaaclab/test/utils/test_tensor_processors.py -q — 11 passed.
• uv run --no-sync python -m pytest source/isaaclab/test/managers/test_manager_base.py -q — 8 passed; PyTorch emitted its existing torch.jit.script deprecation warnings.
• Before the compatibility fix, the four new regression cases failed for the intended reasons: the constructor rejected modifiers, both managers ignored assigned legacy fields without warning, and conflicting fields were accepted.
• uv run --extra dev isaaclab --docs — clean full HTML build with warnings treated as errors.
• uv run isaaclab --format — repository formatting and lint checks.
• uv run python tools/changelog/cli.py check upstream/main --include-worktree — changelog validation.
• git diff --check — clean.

Type of change

• Tensor processor API addition with deprecated modifier aliases
• Unified tensor and structured-data processor interface
• Documentation update

Release backport

• Backport this pull request to the active release branch after it merges into develop

[mataylor-nvidia]

Another option is to do aliasing, however this delays the name change that will eventually come

[mataylor-nvidia]

run-ci

[mataylor-nvidia]

run-ci

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new actionable issues were found.

Summary

Renames modifiers to tensor processors and gives tensor and structured-data processors one shared interface.

• Tensor processors now accept nested data and return structured results.
• Observation terms now configure processors with tensor_processors.
• Old modifier imports still work with deprecation warnings.
• Warp observations run processors on their existing output buffers.

The earlier test-coverage thread was manually resolved without explanation. The replacement interface now has concrete coverage.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Observation configuration] --> B{Legacy modifiers set?}
    B -->|Only legacy field| C[Warn and forward to tensor_processors]
    B -->|Both fields| D[Raise ValueError]
    B -->|No| E[Prepare processors]
    C --> E
    E --> F[Compute observation]
    F --> G[Processors]
    G --> H[Noise, clipping, scaling, delay, history]

Loading

_{Reviews (2) · Last reviewed commit: "Address tensor processor review feedback" · Reviewed by Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

Needs attention: Compatibility and deprecation (1), Style consistency (2). Other reviewed categories look good.

• PR description: The body matches the diff, but replace the fork follow-up link, define PPISP, remove the stale “import-only” and Windows-temp details, and provide exact validation commands.
• Design and architecture: Architecture looks good.
• Scope and over-engineering: No over-engineering issues.
• API: API changes look good.
• Compatibility and deprecation: Breaking changes: ObservationTermCfg.modifiers lacks a functional alias, targeted warning, and transition coverage; retain it as a deprecated forwarding field before migrating consumers.
• Implementation: Implementation looks good.
• Style consistency: Renamed identifiers remain in prose comments; use “tensor processors” outside code references.
• Test quality: Test quality looks good.

Significant concerns. Posted 3 inline findings.

Automated review; human maintainers own approval decisions.

[StafaH]

This doesn't make any sense, transform into the same thing?

[mataylor-nvidia]

updated to single TensorProcessor unified class

[mataylor-nvidia]

@greptile review

[mataylor-nvidia]

run-ci

[hujc7]

The old cfg also need deprecation shim?

[ooctipus]

the input and output key proably only make sense for dict tensor right?