You are generating structured augmentations of a robot task instruction
for training a language-conditioned policy. Unlike free-form rephrasing,
your variants follow a NAMED 5-axis taxonomy — each axis omits or varies
a specific element of the task while preserving its meaning.

Original task: "{base_task}"

Produce variants along five named axes. Each axis has a target count.
The whole batch should expose the policy to maximum linguistic diversity
WITHOUT changing what the robot is supposed to do.

Axes and target counts:

  synonym_paraphrase ({n_synonym}):
    Different wording / verbs / sentence structure. ALL information
    from the original task is preserved — same object, same arm
    specification if present, same orientation if present, same grasp
    if present.

  omit_arm ({n_omit_arm}):
    Drop the left/right/both arm specification from the task. Skip
    entirely (emit 0 entries) if the original task does NOT mention an
    arm. Do not invent an arm specification just to omit it.

  omit_orientation ({n_omit_orientation}):
    Drop orientation cues (upright, sideways, facing the user,
    long-edge-first, etc.). Skip entirely if no orientation cue is
    present in the original task.

  omit_grasp_method ({n_omit_grasp_method}):
    Drop the grip / grasp method specification (pinch, wrap, hold by
    the rim, etc.). Skip entirely if no grasp method is mentioned.

  combined_omissions ({n_combined}):
    Combine TWO of the above omissions simultaneously (e.g. drop both
    arm and orientation). Skip entirely if fewer than two of (arm,
    orientation, grasp_method) appear in the original task.

Hard rules:
- Each variant MUST preserve the core action and the target object.
  Do not change which object is involved, the destination, or the
  high-level action.
- Each variant is plain prose, no markdown, no quotes, no list numbers.
- Each variant must be DISTINCT from every other variant in the entire
  output, both within and across axes. Near-duplicates are not allowed.
- If an axis cannot reach its target count because the original task
  lacks the omittable element, emit fewer entries — do NOT pad the
  axis with paraphrases that belong to a different axis.
- Variants should not all start with verbs — vary sentence structure
  (some imperative, some polite request, some question).

Output strictly valid JSON of shape:

  {{
    "synonym_paraphrase": ["<v1>", "<v2>", ...],
    "omit_arm": ["<v1>", "<v2>", ...],
    "omit_orientation": ["<v1>", ...],
    "omit_grasp_method": ["<v1>", ...],
    "combined_omissions": ["<v1>", ...]
  }}
