rskill-playbook-verify_outcome
A kind: playbook rSkill: a symbolic S2 decision procedure the Reasoner
reads, not a neural policy. It has no weights; the authored
PLAYBOOK.md is its runtime.
What this skill does
Closes the loop after a skill or subtask completes (Inner Monologue). The action
result ("the policy reported done") is not the same as the world state, so this
playbook verifies what actually happened: it asks the scene a specific yes/no
success question, cross-checks a reward monitor, optionally confirms the object's
pose, and classifies success or failure with evidence before the reasoner
proceeds. A failure routes to a replan or human handoff β never a silent
"success". Concrete walkthrough: the black-bowl-on-plate example in
PLAYBOOK.md.
How it works
This playbook is content, not code. When installed, the reasoner injects
PLAYBOOK.md into its system prompt and follows the SOP, composing tools it
already has (query_scene, query_task_progress, locate_in_view,
memory_write, emit_prompt). It is role: s2 and is never dispatched
through ExecuteRskill. Any motion a downstream replan triggers is an
execute_rskill β Action chunk β C++ safety kernel β the playbook holds no
actuation authority (CLAUDE.md Β§1.1).
Observation β action contract
None. A playbook emits no Action chunks and requires no actuators
(actuators_required: [], chunk_size: 1). Its "output" is the sequence of
tool calls the reasoner makes while following the SOP, bounded by
playbook.max_steps.
How it was authored / Upstream provenance
N/A: hand-authored, not trained, so no weights and no upstream model.
Provenance is the authoring decision record (linked via paper_url). To change
behaviour, edit PLAYBOOK.md and bump version.
Supported robots
Embodiment-agnostic β declares the explicit wildcard embodiment_tags: ["any"]
(never an empty list). Gated by capabilities_required
(has_vision: true β a real RobotCapabilities flag): the loader filters it out
on robots without a camera. The scene / reward queries are gated at runtime by
the composed tools, not by this playbook's flags.
Sensors required
None directly. The tools it composes declare their own sensor needs.
Manifest summary
kind: playbook,role: s2,actions: [plan],chunk_size: 1.playbook.trigger: a skill or subtask has just completed and its success is not directly observable from the action result.playbook.done_predicate: the outcome is classified success or failure with evidence, and a failure has triggered a replan or handoff.playbook.max_steps: 6.
Quick start
from openral_core.schemas import RSkillManifest
m = RSkillManifest.from_yaml("rskills/verify-outcome/rskill.yaml")
assert m.kind == "playbook" and m.playbook is not None
print(m.playbook.trigger)
Reproduction
Packaging-only: tests/unit/test_rskill_manifest_kinds.py loads every in-tree
manifest, this one included; the kind: playbook rules and the reasoner's
playbook injection are tested against rskills/find-object
(tests/unit/test_playbook_rskill_manifest.py,
tests/unit/test_reasoner_playbooks.py). No benchmark number to reproduce.
Evaluation
N/A β no eval/*.json; a playbook produces no benchmarkable policy output.
License
- Code / content: Apache-2.0.
- Weights: none.
See also
PLAYBOOK.mdβ the decision procedure itself.