code/test/test_vrbench_navigation_training.py

#!/usr/bin/env python3
"""Dependency-light tests for VRBench SPAgent training environments."""

from __future__ import annotations

import importlib.util
import os
import sys
import tempfile
import types
from pathlib import Path

from PIL import Image


ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))

from spagent.frozenlake_v3_protocol import count_explicit_images  # noqa: E402
from spagent.vrbench_navigation_protocol import (  # noqa: E402
    EMPTY,
    GOAL,
    MAZE_TASK,
    PATHFINDER_TASK,
    TRAP,
    TRAPFIELD_SOLVABLE_SEEDS,
    TRAPFIELD_TASK,
    WALL,
    build_system_prompt,
    build_user_prompt,
    canonical_state,
    encode_initial_state,
    render_state,
    replay_actions,
    shortest_path,
    state_text,
    step_state,
)

_builder_spec = importlib.util.spec_from_file_location(
    "build_vrbench_navigation_training_data",
    ROOT / "train" / "build_vrbench_navigation_training_data.py",
)
builder = importlib.util.module_from_spec(_builder_spec)
assert _builder_spec.loader is not None
_builder_spec.loader.exec_module(builder)


def simple_state() -> dict:
    return canonical_state(
        grid=[
            [WALL, WALL, WALL, WALL, WALL],
            [WALL, EMPTY, EMPTY, EMPTY, WALL],
            [WALL, WALL, WALL, EMPTY, WALL],
            [WALL, EMPTY, EMPTY, GOAL, WALL],
            [WALL, WALL, WALL, WALL, WALL],
        ],
        start=(1, 1),
        goal=(3, 3),
    )


def trap_state() -> dict:
    return canonical_state(
        grid=[
            [WALL, WALL, WALL, WALL, WALL],
            [WALL, EMPTY, TRAP, EMPTY, WALL],
            [WALL, EMPTY, EMPTY, EMPTY, WALL],
            [WALL, EMPTY, EMPTY, GOAL, WALL],
            [WALL, WALL, WALL, WALL, WALL],
        ],
        start=(1, 1),
        goal=(3, 3),
    )


def test_protocol_and_replay() -> None:
    state = simple_state()
    solution = ["Right", "Right", "Down", "Down"]
    assert shortest_path(state) == solution
    assert replay_actions(MAZE_TASK, encode_initial_state(state), solution)
    assert not replay_actions(
        MAZE_TASK,
        encode_initial_state(state),
        ["Down"],
    )


def test_blocked_wall_and_trap() -> None:
    state = simple_state()
    wall = step_state(PATHFINDER_TASK, state, "Down")
    assert not wall.changed
    assert wall.tool_result["error"] == "wall"
    trapped = step_state(TRAPFIELD_TASK, trap_state(), "Right")
    assert not trapped.changed
    assert trapped.tool_result["error"] == "trap"


def test_text_render_and_prompt(tmp: Path) -> None:
    state = simple_state()
    image_path = tmp / "maze.png"
    render_state(MAZE_TASK, state, image_path)
    assert state_text(state, initial=True).splitlines()[1] == "#SFF#"
    prompt = build_user_prompt(MAZE_TASK, state, image_path)
    assert count_explicit_images(prompt) == 1
    assert "# Initial Grid (text view, 5x5)" in prompt
    assert '"name": "move"' in build_system_prompt(MAZE_TASK, state)
    with Image.open(image_path) as image:
        assert image.mode == "RGB"
        assert image.size == (240, 240)


def load_plugin_with_stubs():
    class FakeRolloutInferRequest:
        def __init__(self, **kwargs):
            self.__dict__.update(kwargs)

    class FakeScheduler:
        def __init__(self, max_turns=20, *args, **kwargs):
            self.max_turns = max_turns

        def check_finished(self, infer_request, response_choice, current_turn):
            return current_turn >= self.max_turns

    class FakeORM:
        pass

    class FakeLogger:
        def warning(self, *args, **kwargs):
            pass

    swift = types.ModuleType("swift")
    swift_llm = types.ModuleType("swift.llm")
    swift_llm.RolloutInferRequest = FakeRolloutInferRequest
    swift_infer = types.ModuleType("swift.llm.infer")
    swift_protocol = types.ModuleType("swift.llm.infer.protocol")
    swift_protocol.ChatCompletionResponseChoice = object
    swift_plugin = types.ModuleType("swift.plugin")
    swift_plugin.ORM = FakeORM
    swift_plugin.orms = {}
    swift_multi = types.ModuleType("swift.plugin.multi_turn")
    swift_multi.MultiTurnScheduler = FakeScheduler
    swift_multi.multi_turns = {}
    swift_utils = types.ModuleType("swift.utils")
    swift_utils.get_logger = lambda: FakeLogger()
    trainer = types.ModuleType("grpo_filtered_trainer")

    modules = {
        "swift": swift,
        "swift.llm": swift_llm,
        "swift.llm.infer": swift_infer,
        "swift.llm.infer.protocol": swift_protocol,
        "swift.plugin": swift_plugin,
        "swift.plugin.multi_turn": swift_multi,
        "swift.utils": swift_utils,
        "grpo_filtered_trainer": trainer,
    }
    previous = {name: sys.modules.get(name) for name in modules}
    sys.modules.update(modules)
    try:
        spec = importlib.util.spec_from_file_location(
            "vrbench_navigation_plugin_under_test",
            ROOT / "train" / "vrbench_navigation_plugin.py",
        )
        module = importlib.util.module_from_spec(spec)
        assert spec.loader is not None
        spec.loader.exec_module(module)
        return module, FakeRolloutInferRequest
    finally:
        for name, old in previous.items():
            if old is None:
                sys.modules.pop(name, None)
            else:
                sys.modules[name] = old


def test_reward_and_scheduler(tmp: Path) -> None:
    plugin, request_type = load_plugin_with_stubs()
    state = simple_state()
    encoded = encode_initial_state(state)
    solution = "Right Right Down Down"
    completion = f"<think>Plan.</think><answer>{solution}</answer>"
    assert plugin.VRBenchNavigationReward()(
        [completion],
        [MAZE_TASK],
        [encoded],
    ) == [1.0]
    assert plugin.VRBenchNavigationFormatReward()([completion]) == [0.5]

    initial_image = tmp / "initial.png"
    render_state(MAZE_TASK, state, initial_image)
    old_dir = os.environ.get("VRBENCH_NAV_IMG_DIR")
    os.environ["VRBENCH_NAV_IMG_DIR"] = str(tmp / "rollout")
    try:
        scheduler = plugin.VRBenchNavigationScheduler(max_turns=10)
        request = request_type(
            messages=[
                {"role": "system", "content": "system"},
                {"role": "user", "content": "user"},
            ],
            images=[str(initial_image)],
            audios=None,
            videos=None,
            tools=None,
            objects=None,
            data_dict={
                "task_id": MAZE_TASK,
                "initial_state": encoded,
            },
        )
        choice = types.SimpleNamespace(
            message=types.SimpleNamespace(
                content=(
                    "<think>Move right.</think>"
                    '<tool_call>{"name":"move","arguments":'
                    '{"direction":"Right"}}</tool_call>'
                )
            )
        )
        result = scheduler.step(request, choice, 1)
        next_request = result["infer_request"]
        assert next_request.data_dict["_vrbench_navigation"]["actions"] == [
            "Right"
        ]
        assert len(next_request.images) == 2
        assert Path(next_request.images[-1]).is_file()
        assert count_explicit_images(
            next_request.messages[-1]["content"]
        ) == 1

        blocked_choice = types.SimpleNamespace(
            message=types.SimpleNamespace(
                content=(
                    "<think>Try down.</think>"
                    '<tool_call>{"name":"move","arguments":'
                    '{"direction":"Down"}}</tool_call>'
                )
            )
        )
        blocked = scheduler.step(next_request, blocked_choice, 2)
        blocked_request = blocked["infer_request"]
        assert len(blocked_request.images) == 2
        assert count_explicit_images(
            blocked_request.messages[-1]["content"]
        ) == 1
    finally:
        if old_dir is None:
            os.environ.pop("VRBENCH_NAV_IMG_DIR", None)
        else:
            os.environ["VRBENCH_NAV_IMG_DIR"] = old_dir


def test_trapfield_seed_contract() -> None:
    assert len(TRAPFIELD_SOLVABLE_SEEDS) == 30
    assert len(set(TRAPFIELD_SOLVABLE_SEEDS)) == 30


def main() -> None:
    with tempfile.TemporaryDirectory() as directory:
        tmp = Path(directory)
        test_protocol_and_replay()
        test_blocked_wall_and_trap()
        test_text_render_and_prompt(tmp)
        test_reward_and_scheduler(tmp)
        test_trapfield_seed_contract()
    print("VRBench navigation training tests passed")


if __name__ == "__main__":
    main()