code/spagent/vrbench_navigation_protocol.py

"""Shared SPAgent protocol for deterministic VRBench navigation tasks.

Dataset generation samples official Beyond-MMEnv states. Training rollouts,
answer replay, rendering, and tests use the pure deterministic state machine
defined here, so distributed workers do not need Beyond-MMEnv installed.
"""

from __future__ import annotations

import json
import os
from collections import deque
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, Iterable, List, Mapping, MutableMapping, Tuple

from PIL import Image, ImageDraw, ImageFont

from spagent.frozenlake_v3_protocol import (
    DIRS,
    create_tool_response,
    parse_action_tokens,
)


PATHFINDER_TASK = "VRBench/Pathfinder-v0"
MAZE_TASK = "VRBench/Maze-v0"
TRAPFIELD_TASK = "VRBench/TrapField-v0"
SUPPORTED_TASKS = (PATHFINDER_TASK, MAZE_TASK, TRAPFIELD_TASK)

EMPTY = 0
WALL = 1
PLAYER = 2
GOAL = 3
TRAP = 4

TASK_LABELS = {
    PATHFINDER_TASK: "Pathfinder",
    MAZE_TASK: "Maze",
    TRAPFIELD_TASK: "TrapField",
}
TASK_TITLES = {
    PATHFINDER_TASK: "VRBench Pathfinder",
    MAZE_TASK: "VRBench Maze",
    TRAPFIELD_TASK: "VRBench TrapField",
}

# The current upstream TrapField generator usually disconnects start and goal.
# These seeds are verified solvable and are also used by the evaluation suite.
TRAPFIELD_SOLVABLE_SEEDS = (
    173,
    1126,
    1553,
    1701,
    2596,
    3860,
    3962,
    4296,
    4736,
    5098,
    5206,
    5805,
    5839,
    7126,
    7856,
    10060,
    10741,
    10748,
    10751,
    10995,
    11073,
    11089,
    11155,
    11194,
    11918,
    12206,
    13519,
    14715,
    15212,
    15265,
)

CELL_PX = 40
MARGIN = 20
PALETTE = {
    "F": (210, 235, 255),
    "S": (170, 225, 150),
    "G": (245, 205, 70),
    "#": (45, 55, 70),
    "X": (190, 65, 65),
    "AGENT": (220, 70, 70),
    "GRID": (60, 60, 60),
}
_font: ImageFont.ImageFont | None = None


@dataclass(frozen=True)
class StepOutcome:
    changed: bool
    done: bool
    success: bool
    body_text: str
    tool_result: Dict[str, Any]


def encode_initial_state(state: Mapping[str, Any]) -> str:
    return json.dumps(state, sort_keys=True, separators=(",", ":"))


def decode_initial_state(value: str | Mapping[str, Any]) -> Dict[str, Any]:
    if isinstance(value, str):
        state = json.loads(value)
    elif isinstance(value, Mapping):
        state = dict(value)
    else:
        raise TypeError(
            "initial_state must be JSON text or a mapping, "
            f"got {type(value)}"
        )
    if not isinstance(state, dict):
        raise ValueError("initial_state must decode to an object")
    return state


def clone_state(state: Mapping[str, Any]) -> Dict[str, Any]:
    return json.loads(json.dumps(state))


def canonical_state(
    *,
    grid: Iterable[Iterable[int]],
    start: Tuple[int, int],
    goal: Tuple[int, int],
) -> Dict[str, Any]:
    rows = [[int(cell) for cell in row] for row in grid]
    if not rows or not rows[0] or any(len(row) != len(rows[0]) for row in rows):
        raise ValueError("grid must be a non-empty rectangle")
    allowed = {EMPTY, WALL, PLAYER, GOAL, TRAP}
    if any(cell not in allowed for row in rows for cell in row):
        raise ValueError("grid contains an unsupported cell value")
    sr, sc = (int(start[0]), int(start[1]))
    gr, gc = (int(goal[0]), int(goal[1]))
    if not (0 <= sr < len(rows) and 0 <= sc < len(rows[0])):
        raise ValueError(f"start is outside the grid: {(sr, sc)}")
    if not (0 <= gr < len(rows) and 0 <= gc < len(rows[0])):
        raise ValueError(f"goal is outside the grid: {(gr, gc)}")
    rows[sr][sc] = EMPTY
    rows[gr][gc] = GOAL
    return {
        "grid": rows,
        "start": [sr, sc],
        "goal": [gr, gc],
        "position": [sr, sc],
    }


def symbolic_grid_lines(
    state: Mapping[str, Any],
    *,
    initial: bool = False,
) -> List[str]:
    grid = state["grid"]
    position = tuple(int(value) for value in state["position"])
    start = tuple(int(value) for value in state["start"])
    goal = tuple(int(value) for value in state["goal"])
    mapping = {EMPTY: "F", WALL: "#", GOAL: "G", TRAP: "X"}
    rows: List[str] = []
    for r, row in enumerate(grid):
        chars: List[str] = []
        for c, value in enumerate(row):
            pos = (r, c)
            if initial and pos == start:
                chars.append("S")
            elif not initial and pos == position:
                chars.append("P")
            elif pos == goal:
                chars.append("G")
            else:
                chars.append(mapping[int(value)])
        rows.append("".join(chars))
    return rows


def state_text(state: Mapping[str, Any], *, initial: bool = False) -> str:
    return "\n".join(symbolic_grid_lines(state, initial=initial))


def is_solved(state: Mapping[str, Any]) -> bool:
    position = tuple(int(value) for value in state["position"])
    goal = tuple(int(value) for value in state["goal"])
    return position == goal


def step_state(
    task_id: str,
    state: MutableMapping[str, Any],
    direction: Any,
) -> StepOutcome:
    if task_id not in SUPPORTED_TASKS:
        raise ValueError(f"Unsupported task_id: {task_id}")
    if not isinstance(direction, str) or direction not in DIRS:
        return StepOutcome(
            changed=False,
            done=False,
            success=False,
            body_text=(
                f"Invalid direction {direction!r}. "
                "Use one of: Left, Down, Right, Up."
            ),
            tool_result={
                "success": False,
                "error": "invalid_direction",
                "direction": direction,
            },
        )

    grid = state["grid"]
    r, c = (int(value) for value in state["position"])
    dr, dc = DIRS[direction]
    nr, nc = r + dr, c + dc
    if not (0 <= nr < len(grid) and 0 <= nc < len(grid[0])):
        obstacle = "wall"
    else:
        destination = int(grid[nr][nc])
        obstacle = (
            "wall"
            if destination == WALL
            else "trap"
            if destination == TRAP
            else ""
        )
    if obstacle:
        return StepOutcome(
            changed=False,
            done=False,
            success=False,
            body_text=(
                f"Blocked by {obstacle}: cannot move {direction} from "
                f"({r},{c}). Position unchanged."
            ),
            tool_result={
                "success": False,
                "error": obstacle,
                "from": [r, c],
                "direction": direction,
            },
        )

    state["position"] = [nr, nc]
    success = (nr, nc) == tuple(int(value) for value in state["goal"])
    cell = "G" if success else "F"
    body = f"Moved {direction} to ({nr},{nc}). Cell: {cell}."
    if success:
        body += " **Reached goal - success!**"
    return StepOutcome(
        changed=True,
        done=success,
        success=success,
        body_text=body,
        tool_result={
            "success": True,
            "to": [nr, nc],
            "cell": cell,
            "done": success,
            "succeeded": success,
        },
    )


def replay_actions(
    task_id: str,
    initial_state: str | Mapping[str, Any],
    actions: Iterable[Any],
) -> bool:
    state = clone_state(decode_initial_state(initial_state))
    if is_solved(state):
        return True
    for direction in actions:
        outcome = step_state(task_id, state, direction)
        if not outcome.changed:
            return False
        if outcome.success:
            return True
    return is_solved(state)


def shortest_path(state: Mapping[str, Any]) -> List[str] | None:
    grid = state["grid"]
    start = tuple(int(value) for value in state["start"])
    goal = tuple(int(value) for value in state["goal"])
    queue = deque([(start, [])])
    seen = {start}
    while queue:
        (r, c), path = queue.popleft()
        if (r, c) == goal:
            return path
        for direction, (dr, dc) in DIRS.items():
            nr, nc = r + dr, c + dc
            if not (0 <= nr < len(grid) and 0 <= nc < len(grid[0])):
                continue
            if (nr, nc) in seen or int(grid[nr][nc]) in (WALL, TRAP):
                continue
            seen.add((nr, nc))
            queue.append(((nr, nc), path + [direction]))
    return None


def _font_fn() -> ImageFont.ImageFont:
    global _font
    if _font is None:
        try:
            _font = ImageFont.truetype("DejaVuSans-Bold.ttf", 18)
        except OSError:
            _font = ImageFont.load_default()
    return _font


def render_state(
    task_id: str,
    state: Mapping[str, Any],
    output_path: str | Path,
) -> None:
    if task_id not in SUPPORTED_TASKS:
        raise ValueError(f"Unsupported task_id: {task_id}")
    lines = symbolic_grid_lines(state, initial=True)
    rows, cols = len(lines), len(lines[0])
    width = cols * CELL_PX + 2 * MARGIN
    height = rows * CELL_PX + 2 * MARGIN
    image = Image.new("RGB", (width, height), "white")
    draw = ImageDraw.Draw(image)
    font = _font_fn()
    for row_index, row in enumerate(lines):
        for col_index, value in enumerate(row):
            x0 = MARGIN + col_index * CELL_PX
            y0 = MARGIN + row_index * CELL_PX
            draw.rectangle(
                [x0, y0, x0 + CELL_PX, y0 + CELL_PX],
                fill=PALETTE[value],
                outline=PALETTE["GRID"],
                width=2,
            )
            bbox = draw.textbbox((0, 0), value, font=font)
            text_width = bbox[2] - bbox[0]
            draw.text(
                (
                    x0 + (CELL_PX - text_width) // 2,
                    y0 + 2,
                ),
                value,
                fill="white" if value in {"#", "X"} else "black",
                font=font,
            )
    r, c = (int(value) for value in state["position"])
    center_x = MARGIN + c * CELL_PX + CELL_PX // 2
    center_y = MARGIN + r * CELL_PX + CELL_PX // 2 + 5
    radius = max(6, CELL_PX // 4)
    draw.ellipse(
        [
            center_x - radius,
            center_y - radius,
            center_x + radius,
            center_y + radius,
        ],
        fill=PALETTE["AGENT"],
        outline="black",
        width=2,
    )
    mode = os.environ.get("TOOL_IMAGE_MODE", "real").lower().strip()
    if mode == "black":
        image = Image.new("RGB", image.size, "black")
    elif mode == "white":
        image = Image.new("RGB", image.size, "white")
    output_path = Path(output_path)
    output_path.parent.mkdir(parents=True, exist_ok=True)
    image.save(output_path, format="PNG")


def move_tool_spec(task_id: str) -> Dict[str, Any]:
    label = TASK_LABELS[task_id]
    blocked = "walls and blocked traps" if task_id == TRAPFIELD_TASK else "walls"
    return {
        "type": "function",
        "function": {
            "name": "move",
            "description": (
                f"Move the agent one cell on the {label} grid. Moves into "
                f"{blocked} are no-ops. Returns the updated complete grid "
                "as a new image."
            ),
            "parameters": {
                "type": "object",
                "properties": {
                    "direction": {
                        "type": "string",
                        "enum": ["Left", "Down", "Right", "Up"],
                    }
                },
                "required": ["direction"],
            },
        },
    }


def build_system_prompt(task_id: str, state: Mapping[str, Any]) -> str:
    if task_id not in SUPPORTED_TASKS:
        raise ValueError(f"Unsupported task_id: {task_id}")
    size = len(state["grid"])
    label = TASK_LABELS[task_id]
    cells = (
        "S = start, G = goal, F = free cell, # = wall, X = blocked trap."
        if task_id == TRAPFIELD_TASK
        else "S = start, G = goal, F = free cell, # = wall."
    )
    objective = (
        "Reach G while avoiding walls and every blocked trap X."
        if task_id == TRAPFIELD_TASK
        else "Reach G while avoiding walls."
    )
    schema = json.dumps(move_tool_spec(task_id), ensure_ascii=False)
    return (
        f"You are an agent navigating a {size}x{size} {label} grid.\n\n"
        f"Cells: {cells}\n"
        "The environment is deterministic and fully observed. A move succeeds "
        "when its destination is traversable; blocked moves are no-ops and the "
        "grid does not wrap.\n"
        f"Goal: {objective}\n\n"
        "# Tools\n\n"
        "You may call one or more functions to assist with the user query.\n\n"
        "You are provided with function signatures within <tools></tools> XML "
        "tags:\n"
        "<tools>\n"
        f"{schema}\n"
        "</tools>\n\n"
        "For each function call, return a JSON object with function name and "
        "arguments within <tool_call></tool_call> XML tags:\n"
        "<tool_call>\n"
        '{"name": <function-name>, "arguments": <args-json-object>}\n'
        "</tool_call>\n\n"
        "Reason step-by-step in <think></think>. When you reach G, output the "
        "full executed direction sequence in <answer></answer>, space-separated."
    )


def build_user_prompt(
    task_id: str,
    state: Mapping[str, Any],
    image_path: str | Path,
) -> str:
    label = TASK_LABELS[task_id]
    lines = symbolic_grid_lines(state, initial=True)
    start = state["start"]
    goal = state["goal"]
    objective = (
        "Avoid walls and every blocked trap X."
        if task_id == TRAPFIELD_TASK
        else "Avoid every wall #."
    )
    return (
        f"# Initial Grid (text view, {len(lines)}x{len(lines[0])})\n"
        + "\n".join(lines)
        + "\n\n# Images\n"
        f"- Image_1 (initial {label} grid) — img_idx=0\n"
        f"<image>[image: {Path(image_path).resolve()}]\n\n"
        f"You are at S (row {start[0]}, col {start[1]}). Reach G at "
        f"(row {goal[0]}, col {goal[1]}). {objective} Call the `move` tool "
        "one step at a time. After each valid move you will see the updated "
        "complete grid.\n\n"
        "Output step-by-step reasoning in <think></think>. When you reach G, "
        "output the full executed action sequence in <answer></answer>."
    )


__all__ = [
    "DIRS",
    "EMPTY",
    "GOAL",
    "MAZE_TASK",
    "PATHFINDER_TASK",
    "PLAYER",
    "SUPPORTED_TASKS",
    "StepOutcome",
    "TRAP",
    "TRAPFIELD_SOLVABLE_SEEDS",
    "TRAPFIELD_TASK",
    "WALL",
    "build_system_prompt",
    "build_user_prompt",
    "canonical_state",
    "clone_state",
    "create_tool_response",
    "decode_initial_state",
    "encode_initial_state",
    "is_solved",
    "parse_action_tokens",
    "render_state",
    "replay_actions",
    "shortest_path",
    "state_text",
    "step_state",
    "symbolic_grid_lines",
]