Fix replay crash on trailing idle action in IK tasks

[rwiltz]

Description

Replaying demonstrations for an absolute task-space (IK) task crashed after all
episodes had replayed successfully, with a torch.linalg.solve ... input matrix is singular error from the differential IK controller instead of exiting cleanly.

Root cause is a trailing step past the end of the recorded data. In
replay_episodes_loop, env.step(actions) sat outside the has_next_action
check, so once every environment had exhausted its episodes the loop applied the
untouched idle_action one last time before terminating. No task defines
idle_action, so it falls back to torch.zeros(env.action_space.shape) — and for
an absolute-pose action ([pos_xyz, quat_xyzw, gripper]) a zeros action carries a
zero-norm quaternion.

DifferentialIKController.set_command renormalized it as quat / norm(quat),
i.e. 0/0, producing a NaN target orientation. The NaN propagated into the joint
position targets, diverged the articulation, and only surfaced on the next
decimation sub-step as an unrelated "singular matrix" failure — which is why the
traceback points at the solver rather than at the defect.

Fixes:

• scripts/tools/replay_demos.py: stop before stepping once every environment is
  exhausted, so the replay terminates without applying the idle action or running
  another IK solve.
• DifferentialIKController.set_command: a degenerate (zero-norm) commanded
  quaternion now holds the current end-effector orientation (identity when none was
  supplied) instead of emitting NaN. Applied per-environment via a branchless
  torch.where, so there is no added host sync on the training hot path.
• adaptive_dls: a non-finite Jacobian is reported by its actual cause instead of
  the opaque LAPACK singular-matrix / convergence failure. The check runs only on
  the failure path, so the happy path is unchanged.

The controller change is not redundant with the script change: with --num_envs > 1,
an environment that finishes early keeps receiving the zero-quaternion idle action
every step while the others replay, which would diverge that environment and take
down the whole batch.

Reproduced and verified with IsaacContrib-Stack-Cube-SO101-IK-Abs-v0. On the
unfixed code the QA command produces a byte-for-byte identical traceback; with the
fix it prints Finished replaying 1 episode. and exits 0.

Note (out of scope): idle_action defaulting to zeros is unsound for any
absolute-pose task, since a zero quaternion is never a valid orientation. Seeding
the quaternion to identity would be a broader change affecting every task using the
replay script, so it is left for a separate PR.

Fixes # (issue)

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (do not edit CHANGELOG.rst or bump extension.toml — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[greptile-apps]

Greptile Summary

This PR prevents demonstration replay from applying a trailing idle action and makes absolute-pose IK commands robust to degenerate quaternions.

• Stops the replay loop before stepping after all environments exhaust their recorded actions.
• Falls back to the current end-effector orientation, or identity, when an absolute-pose quaternion cannot be normalized.
• Adds regression coverage for replay termination and quaternion fallback behavior.

Confidence Score: 4/5

The PR is not yet safe to merge because the previously reported adaptive-DLS diagnostic bypass remains unresolved.

The adaptive-DLS path still sends non-finite Jacobians through svdvals and solve without unconditional input or output validation, allowing backends that propagate non-finite values without raising to return invalid joint targets.

Files Needing Attention: source/isaaclab/isaaclab/controllers/differential_ik.py

Important Files Changed

Filename	Overview
scripts/tools/replay_demos.py	Stops replay before the final idle-action step once all environments are exhausted.
source/isaaclab/isaaclab/controllers/differential_ik.py	Adds per-environment fallback handling for unnormalizable absolute-pose quaternions.
source/isaaclab/test/cli/test_replay_demos_loop_termination.py	Adds regression coverage proving replay performs exactly one step per recorded action.
source/isaaclab/test/controllers/test_differential_ik_features.py	Covers zero, underflowing, and tiny-normalizable absolute-pose quaternions.

_{Reviews (3): Last reviewed commit: "Drop adaptive DLS Jacobian guard and tig..." | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The replay loop now avoids the trailing idle step, and the IK controller prevents zero-quaternion commands from producing NaNs. Two API details need correction before merge: the tolerance changes behavior for tiny nonzero quaternions, and the new absolute-pose fallback semantics are absent from the public contract.

• Design and architecture: The replay termination change preserves episode bookkeeping while removing the erroneous final step. The per-environment orientation fallback is appropriately localized, but its threshold should not broaden the fix beyond genuinely degenerate commands without an intentional compatibility decision.
• API: set_command previously normalized every nonzero quaternion, whereas quaternions below 1e-6 now select a fallback orientation. Additionally, ee_quat now influences degenerate absolute-pose commands despite the docstring describing it as relevant only to position and relative-pose modes. Narrow the degeneracy condition or document the compatibility change, and document the new fallback behavior.
• Implementation: The adaptive-DLS exception handling preserves the original exception for finite Jacobians and provides a clearer error for non-finite inputs. Tests cover zero-quaternion identity, current-orientation, and per-environment fallback paths, but do not cover compatibility for tiny nonzero quaternions. The required Isaac Lab changelog fragment is present.

Minor fixes needed. Posted 2 actionable findings inline.

Automated review; human maintainers own approval decisions.

[rwiltz]

@greptile

[AntoineRichard]

Looks reasonable to me, take a look at the other comments.

[AntoineRichard]

AI-generated review — changes requested.

The replay termination and quaternion fallback behave correctly in the focused tests, but this needs another cleanup pass for showroom quality:

• Add direct regression coverage for the primary replay-loop fix. The controller tests do not prove that the extra idle step stays removed.
• Remove or redesign the unconditional adaptive-DLS finiteness diagnostic. Converting the device reduction to a Python boolean synchronizes the host on every controller step, and its comments/error overstate what can be inferred from a non-finite Jacobian.
• Keep comments, public documentation, tests, and the changelog functional and concise. The same incident narrative is currently repeated at every layer.
• Consolidate overlapping quaternion tests and remove the finite-Jacobian guard test already covered by the existing adaptive-DLS behavior test.

Verification: ./isaaclab.sh -p -m pytest source/isaaclab/test/controllers/test_differential_ik_features.py -q passed (21 tests) on the PR head; the targeted regression cases failed on the exact base commit; and ./isaaclab.sh -f passed. The three red CI jobs are unrelated rendering/numerical flakes or a task-environment process crash; all core IsaacLab shards passed.

[rwiltz]

@greptile