Register RLinf GR00T obs/action converters for N1.6/N1.7 + config-driven mapping

[johnnynunez]

Description

The isaaclab_contrib RLinf GR00T integration (source/isaaclab_contrib/isaaclab_contrib/rl/rlinf/extension.py) registered its IsaacLab↔GR00T obs/action converters only on rlinf.models.embodiment.gr00t (GR00T N1.5). RLinf also ships gr00t_n1d6 (N1.6) and gr00t_n1d7 (N1.7, Cosmos‑Reason2‑2B / Qwen3‑VL) embodiment modules, each with its own simulation_io.OBS_CONVERSION / ACTION_CONVERSION table — so GR00T 1.6/1.7 RL never received the IsaacLab converters and failed to find obs_converter_type at runtime.

This PR makes the converter registration GR00T‑version aware and the converters config‑driven:

• _register_gr00t_converters now registers the converters on every available GR00T embodiment module — gr00t, gr00t_n1d6, gr00t_n1d7 — each guarded by a try/except import, instead of only gr00t.
• _convert_isaaclab_obs_to_gr00t supports optional per‑state‑group scale/offset (unit conversion between sim and the checkpoint's training units) and a configurable language_key (N1.6/N1.7 checkpoints commonly use annotation.human.task_description rather than the LIBERO annotation.human.action.task_description).
• _convert_gr00t_to_isaaclab_action honors a configured gr00t_action_keys ordering and optional action scale/offset, and accepts decoded action keys with or without the action. prefix.

All of this is read from the existing env.train.isaaclab.gr00t_mapping / action_mapping config blocks, so no per‑robot Python converter is needed.

Type of change

• New feature (non‑breaking change which adds functionality)

This is additive and backward‑compatible: defaults reproduce the prior N1.5 behavior, and the N1.6/N1.7 registrations are silent no‑ops when those RLinf modules are not installed.

Screenshots / testing

Verified end‑to‑end with a GR00T N1.7 (Cosmos‑Reason2‑2B) PPO run on a custom SO‑101 manipulation task via scripts/reinforcement_learning/train.py --rl_library rlinf (FSDP actor + HuggingFace rollout): multi‑env (4) and multi‑batch (global_batch_size=4) training, 50 epochs, finite advantages / policy loss / grad norm throughout, value loss decreasing. The same path also exercises gr00t/gr00t_n1d6 unchanged.

Note: the gr00t_n1d7 registration becomes active once RLinf's GR00T N1.7 (Gr00tN1d7) embodiment integration is present (proposed upstream to RLinf separately); until then it is skipped gracefully.

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format (formatting unchanged; edit is import/branch logic)
• I have made corresponding changes to the documentation (n/a — internal converter wiring)
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works (manual e2e RL run; no unit-test harness exists for the RLinf converter path)
• I have updated the changelog (maintainer guidance welcome)

[isaaclab-review-bot]

🤖 Isaac Lab Review Bot

PR #5873: Register RLinf GR00T obs/action converters for N1.6/N1.7 + config-driven mapping

✅ Summary

This PR extends the RLinf GR00T integration to support N1.6 and N1.7 embodiment modules (in addition to the existing N1.5 support) and adds config-driven unit conversion capabilities. The changes are additive and backward-compatible.

---

📋 Code Review

_register_gr00t_converters (Registration Logic)

• ✅ Good: Gracefully handles missing modules via try/except
• ✅ Good: Registers converters to all available GR00T versions
• ✅ Good: Clear, informative logging with module names

Minor suggestion: Consider using except ImportError instead of except Exception for the module imports (lines 268, 275, 282). This is more precise and avoids accidentally catching unrelated exceptions during import:

except ImportError as exc:
    logger.debug(f\"Could not import GR00T N1.5 simulation_io: {exc}\")

_convert_isaaclab_obs_to_gr00t (Observation Conversion)

• ✅ Good: Optional scale/offset support enables unit conversions between sim and checkpoint training units
• ✅ Good: Configurable language_key accommodates different checkpoint naming conventions
• ✅ Good: Backward compatible - defaults to annotation.human.action.task_description

_convert_gr00t_to_isaaclab_action (Action Conversion)

• ✅ Good: Configurable gr00t_action_keys ordering provides flexibility
• ✅ Good: Handles both action.* prefixed and unprefixed keys robustly
• ✅ Good: Clear KeyError message when expected keys are missing
• ✅ Good: Optional scale/offset for action space transformations

---

🔍 Observations

1. Backward Compatibility: All defaults preserve the prior N1.5 behavior ✅
2. Graceful Degradation: N1.6/N1.7 registrations are silent no-ops when modules are absent ✅
3. Testing: Author verified with end-to-end GR00T N1.7 PPO training (50 epochs, stable losses)
4. No new warnings: Pre-commit checks pass

---

📊 Verdict

LGTM 👍 - Clean, well-documented feature addition with proper error handling and backward compatibility. The minor suggestion above is non-blocking.

---

Automated review by Isaac Lab Review Bot 🦾

---

Update (commit 752f047): New commits add --rl_model_path CLI support, RL-finetuned weight loading, a warning for missing action keys, and extensive documentation/test updates unrelated to this PR's core feature.

⚠️ Previous suggestion (use except ImportError): Not addressed — non-blocking, original comment stands.

🔴 New issue in extension.py → _convert_gr00t_to_isaaclab_action: The padding/scale/offset block appears corrupted — there is an incomplete np.pad( call (no closing parenthesis or arguments) immediately followed by the scale/offset logic, then the original padding block is duplicated below. This will likely cause a SyntaxError or produce incorrect action transformations. The intended order (pad → scale → offset) needs to be restored as a single coherent block.

---

Update (commit 4d238ab):

✅ Fixed: The broken/duplicate padding block in _convert_gr00t_to_isaaclab_action has been corrected — np.pad now has proper arguments and the pad → scale → offset ordering is clean and coherent.

⚠️ Still open (non-blocking): except ImportError suggestion — original comment stands.

No new issues introduced in this commit.

[greptile-apps]

Greptile Summary

This PR extends the IsaacLab\u2194GR00T converter registration to cover N1.6 and N1.7 embodiment modules (previously only N1.5 was registered), and makes the obs/action converters config-driven with optional scale/offset unit conversion and a configurable language_key.

• _register_gr00t_converters now iterates over all three GR00T simulation-IO modules, each guarded by a try/except so missing versions are silently skipped.
• _convert_isaaclab_obs_to_gr00t reads language_key and per-state-group scale/offset from the YAML config; _convert_gr00t_to_isaaclab_action adds ordered gr00t_action_keys resolution, key-prefix normalisation, and action-level scale/offset.

Confidence Score: 3/5

The action converter has two logic defects in newly added code that would produce wrong robot actions when both padding and offset, or a partially-matching key list, are configured.

The action converter applies scale/offset after zero-padding, meaning any non-zero offset corrupts the padded joint positions. Separately, a configured gr00t_action_keys entry absent from the action chunk is silently dropped, shrinking the output tensor without any warning. Both defects are on the hot path for every action step.

source/isaaclab_contrib/isaaclab_contrib/rl/rlinf/extension.py — specifically the action-converter transform ordering and the missing-key handling in _convert_gr00t_to_isaaclab_action.

Important Files Changed

Filename	Overview
source/isaaclab_contrib/isaaclab_contrib/rl/rlinf/extension.py	Extends GR00T converter registration to N1.6/N1.7 with config-driven scale/offset and action key ordering; has two logic bugs: scale/offset applied after zero-padding (corrupts padded dims when offset≠0) and silent dropping of individual missing action keys.

Comments Outside Diff (1)

1. source/isaaclab_contrib/isaaclab_contrib/rl/rlinf/extension.py, line 53 (link)

   Module-level hard import of GR00T N1.5

   embodiment_tags is unconditionally imported from rlinf.models.embodiment.gr00t (N1.5) at module load time. The three simulation_io imports inside _register_gr00t_converters are now guarded with try/except, but this top-level import is not. If a user has only N1.6 or N1.7 installed without the N1.5 package, the entire extension fails to import — silently negating the graceful-fallback story of the rest of the PR. Wrapping this in a try/except (or moving it inside _patch_embodiment_tags / _patch_gr00t_get_model where it is actually consumed) would make the behavior consistent.

_{Reviews (1): Last reviewed commit: "Merge branch 'develop' into feature/rlin..." | Re-trigger Greptile}