File size: 6,825 Bytes
20c251e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
from __future__ import annotations

from dovla_cil.data.schema import ActionChunk, FailureInfo, RewardInfo, StructuredEffect
from dovla_cil.tasks.library import ToyTaskLibrary
from dovla_cil.vlm.annotation import VLMFailureAnnotator
from dovla_cil.vlm.client import VLMClient


def test_mock_vlm_annotation_works(monkeypatch, tmp_path) -> None:
    monkeypatch.setenv("OPENCLAUDE_MOCK", "1")
    task = ToyTaskLibrary().get_by_id("toy_pick_object_among_distractors")
    local_failure = _failure("wrong_target")
    annotator = VLMFailureAnnotator(
        client=VLMClient(api_key="test-secret", model="mock-model"),
        cache_path=tmp_path / "annotations.json",
    )

    annotated = annotator.annotate_failure(
        task=task,
        instruction=task.instruction,
        action=_action(candidate_type="wrong_target", target="blue_mug"),
        effect=_effect(),
        reward=_reward(success=False),
        local_failure=local_failure,
    )

    assert annotated.type == "wrong_target"
    assert "Mock VLM explanation" in (annotated.language_explanation or "")
    metadata = annotated.metadata["semantic_annotation"]["vlm_annotation"]
    assert metadata["source"] == "vlm"
    assert metadata["suggested_failure_type"] == "wrong_target"
    assert "test-secret" not in str(annotated.to_dict())


def test_annotation_cache_hit_avoids_client_call(tmp_path) -> None:
    task = ToyTaskLibrary().get_by_id("toy_lift_can")
    client = CountingAnnotationClient()
    annotator = VLMFailureAnnotator(client=client, cache_path=tmp_path / "cache.json")
    kwargs = {
        "task": task,
        "instruction": task.instruction,
        "action": _action(candidate_type="no_motion", target="can"),
        "effect": _effect(),
        "reward": _reward(success=False),
        "local_failure": _failure("no_motion"),
    }

    first = annotator.annotate_failure(**kwargs)
    second = annotator.annotate_failure(**kwargs)

    assert client.calls == 1
    assert first.language_explanation == second.language_explanation
    assert second.metadata["semantic_annotation"]["vlm_annotation"]["cache_hit"] is True


def test_invalid_vlm_json_falls_back_to_local_explanation(tmp_path) -> None:
    task = ToyTaskLibrary().get_by_id("toy_lift_can")
    local_failure = _failure("no_motion")
    annotator = VLMFailureAnnotator(client=InvalidAnnotationClient(), cache_path=tmp_path / "cache.json")

    annotated = annotator.annotate_failure(
        task=task,
        instruction=task.instruction,
        action=_action(candidate_type="noop", target="can"),
        effect=_effect(),
        reward=_reward(success=False),
        local_failure=local_failure,
    )

    assert annotated.type == local_failure.type
    assert annotated.language_explanation == local_failure.language_explanation
    metadata = annotated.metadata["semantic_annotation"]["vlm_annotation"]
    assert metadata["source"] == "local_fallback"
    assert metadata["fallback_error"]


def test_vlm_annotation_cannot_change_reward_or_local_failure_type(tmp_path) -> None:
    task = ToyTaskLibrary().get_by_id("toy_pick_object_among_distractors")
    reward = _reward(success=False)
    local_failure = _failure("wrong_target")
    annotator = VLMFailureAnnotator(client=OverridingAnnotationClient(), cache_path=tmp_path / "cache.json")

    annotated = annotator.annotate_failure(
        task=task,
        instruction=task.instruction,
        action=_action(candidate_type="wrong_target", target="blue_mug"),
        effect=_effect(),
        reward=reward,
        local_failure=local_failure,
    )

    assert reward.progress == 0.25
    assert reward.success is False
    assert reward.terminal_success is False
    assert annotated.type == "wrong_target"
    assert annotated.metadata["semantic_annotation"]["suggested_failure_type"] == "success"


class CountingAnnotationClient:
    def __init__(self) -> None:
        self.calls = 0

    def chat_json(self, system: str, user: str, schema_hint=None) -> dict:
        del system, user, schema_hint
        self.calls += 1
        return {
            "failure_type": "no_motion",
            "explanation": "The action produced no task-relevant motion.",
            "avoidance_hint": "Choose an action that contacts the can.",
            "confidence": 0.9,
        }


class InvalidAnnotationClient:
    def chat_json(self, system: str, user: str, schema_hint=None) -> dict:
        del system, user, schema_hint
        return {"failure_type": "no_motion", "confidence": "high"}


class OverridingAnnotationClient:
    def chat_json(self, system: str, user: str, schema_hint=None) -> dict:
        del system, user, schema_hint
        return {
            "failure_type": "success",
            "explanation": "The VLM incorrectly claims this succeeded.",
            "avoidance_hint": "No hint.",
            "confidence": 1.0,
        }


def _action(*, candidate_type: str, target: str) -> ActionChunk:
    return ActionChunk(
        representation="semantic",
        values=[{"command": "grasp", "object": target}],
        skill_type="grasp",
        metadata={
            "candidate_type": candidate_type,
            "intended_target": target,
            "intended_relation": "grasped",
            "difficulty": 0.3,
        },
    )


def _effect() -> StructuredEffect:
    before = {
        "objects": {
            "red_mug": {"position": [0.0, 0.0, 0.03], "grasped": False, "lifted": False},
            "blue_mug": {"position": [0.4, 0.0, 0.03], "grasped": False, "lifted": False},
        },
        "robot": {"eef_position": [0.0, 0.0, 0.2], "gripper": "open", "held_object": None},
    }
    after = {
        "objects": {
            "red_mug": {"position": [0.0, 0.0, 0.03], "grasped": False, "lifted": False},
            "blue_mug": {"position": [0.6, 0.0, 0.03], "grasped": False, "lifted": False},
        },
        "robot": {"eef_position": [0.4, 0.0, 0.2], "gripper": "closed", "held_object": None},
    }
    return StructuredEffect(
        object_pose_delta={"blue_mug": [0.2, 0.0, 0.0]},
        contact_events=[{"type": "touch", "object": "blue_mug"}],
        relation_before={"grasped(red_mug)": False},
        relation_after={"grasped(red_mug)": False},
        grasp_success=False,
        moved_objects=["blue_mug"],
        symbolic_before=before,
        symbolic_after=after,
    )


def _reward(*, success: bool) -> RewardInfo:
    return RewardInfo(
        progress=1.0 if success else 0.25,
        success=success,
        terminal_success=success,
        dense_components={"partial": 0.25},
    )


def _failure(failure_type: str) -> FailureInfo:
    return FailureInfo(
        type=failure_type,
        symbolic_reason=f"local {failure_type}",
        language_explanation=f"Local explanation for {failure_type}.",
    )