gpudrive-main/data_utils/nuscenes_adapter.py

"""NuScenes to GPUDrive JSON conversion utilities.

The adapter keeps NuScenes-specific logic at the dataset boundary and emits the
same JSON schema consumed by GPUDrive's existing C++ map reader.
"""

from __future__ import annotations

from collections import defaultdict
from dataclasses import dataclass
import json
import logging
import math
from pathlib import Path
from typing import Any, Iterable, Optional

import numpy as np

try:
    from .datatypes import MapElementIds
except ImportError:  # pragma: no cover - supports running as a script.
    from datatypes import MapElementIds


ERR_VAL = -1e4


@dataclass
class NuScenesAdapterConfig:
    """Configuration for converting NuScenes scenes into GPUDrive scenarios."""

    target_hz: float = 10.0
    num_steps: int = 91
    max_annotation_gap_sec: float = 1.1
    min_valid_steps: int = 2
    require_start_valid: bool = True
    require_full_window: bool = True
    include_ego: bool = True
    ego_width: float = 2.0
    ego_length: float = 4.8
    ego_height: float = 1.8
    include_map: bool = True
    fail_on_missing_map: bool = False
    map_radius_m: float = 120.0
    map_discretization_m: float = 1.0
    max_road_elements: int = 900
    max_geometry_points: int = 512

    @property
    def dt(self) -> float:
        return 1.0 / self.target_hz

    @property
    def horizon_sec(self) -> float:
        return (self.num_steps - 1) * self.dt


def load_nuscenes(version: str, dataroot: str, verbose: bool = True):
    """Load the NuScenes SDK object with a clear optional-dependency error."""

    try:
        from nuscenes.nuscenes import NuScenes
    except ImportError as exc:  # pragma: no cover - depends on local install.
        raise ImportError(
            "NuScenes conversion requires the optional NuScenes devkit. "
            "Install it with `pip install nuscenes-devkit`."
        ) from exc

    return NuScenes(version=version, dataroot=dataroot, verbose=verbose)


def wrap_yaw(yaw: float | np.ndarray) -> float | np.ndarray:
    """Wrap yaw angles to [-pi, pi]."""

    return (yaw + np.pi) % (2 * np.pi) - np.pi


def yaw_from_quaternion(rotation: Iterable[float]) -> float:
    """Return z-yaw from a NuScenes quaternion stored as [w, x, y, z]."""

    w, x, y, z = rotation
    siny_cosp = 2.0 * (w * z + x * y)
    cosy_cosp = 1.0 - 2.0 * (y * y + z * z)
    return wrap_yaw(math.atan2(siny_cosp, cosy_cosp))


def category_to_gpudrive(category_name: str) -> Optional[str]:
    """Map NuScenes categories to GPUDrive object types."""

    if category_name.startswith("pedestrian."):
        return "pedestrian"
    if category_name in {"vehicle.bicycle", "vehicle.motorcycle"}:
        return "cyclist"
    if category_name.startswith("vehicle."):
        return "vehicle"
    return None


def _dedupe_states_by_time(states: list[dict[str, Any]]) -> list[dict[str, Any]]:
    deduped: dict[float, dict[str, Any]] = {}
    for state in states:
        deduped[float(state["time_sec"])] = state
    return [deduped[t] for t in sorted(deduped)]


def _empty_series(num_steps: int) -> tuple[list[dict[str, float]], list[float], list[dict[str, float]], list[bool]]:
    positions = [{"x": ERR_VAL, "y": ERR_VAL, "z": ERR_VAL} for _ in range(num_steps)]
    headings = [ERR_VAL for _ in range(num_steps)]
    velocities = [{"x": ERR_VAL, "y": ERR_VAL} for _ in range(num_steps)]
    valid = [False for _ in range(num_steps)]
    return positions, headings, velocities, valid


def _interpolate_value(
    raw_t: np.ndarray,
    raw_values: np.ndarray,
    target_t: float,
    right_idx: int,
) -> Optional[np.ndarray]:
    if right_idx < len(raw_t) and abs(raw_t[right_idx] - target_t) <= 1e-6:
        return raw_values[right_idx]
    if right_idx == 0 or right_idx >= len(raw_t):
        return None

    left_idx = right_idx - 1
    denom = raw_t[right_idx] - raw_t[left_idx]
    if denom <= 0:
        return None

    alpha = (target_t - raw_t[left_idx]) / denom
    return raw_values[left_idx] * (1.0 - alpha) + raw_values[right_idx] * alpha


def _is_interpolation_gap_valid(
    raw_t: np.ndarray,
    target_t: float,
    right_idx: int,
    max_gap_sec: float,
) -> bool:
    if right_idx < len(raw_t) and abs(raw_t[right_idx] - target_t) <= 1e-6:
        return True
    if right_idx == 0 or right_idx >= len(raw_t):
        return False
    return raw_t[right_idx] - raw_t[right_idx - 1] <= max_gap_sec


def _compute_velocities(
    xy: np.ndarray,
    valid: np.ndarray,
    dt: float,
) -> list[dict[str, float]]:
    velocities = [{"x": ERR_VAL, "y": ERR_VAL} for _ in range(len(valid))]
    valid_indices = np.flatnonzero(valid)

    for idx in valid_indices:
        prev_valid = idx > 0 and valid[idx - 1]
        next_valid = idx + 1 < len(valid) and valid[idx + 1]
        if prev_valid and next_valid:
            vel = (xy[idx + 1] - xy[idx - 1]) / (2.0 * dt)
        elif next_valid:
            vel = (xy[idx + 1] - xy[idx]) / dt
        elif prev_valid:
            vel = (xy[idx] - xy[idx - 1]) / dt
        else:
            vel = np.array([0.0, 0.0])

        velocities[idx] = {"x": float(vel[0]), "y": float(vel[1])}

    return velocities


def interpolate_track(
    states: list[dict[str, Any]],
    target_times_sec: np.ndarray,
    config: NuScenesAdapterConfig,
) -> Optional[dict[str, Any]]:
    """Interpolate one NuScenes track to GPUDrive's fixed time grid."""

    states = _dedupe_states_by_time(states)
    num_steps = len(target_times_sec)
    if not states:
        return None

    raw_t = np.array([float(s["time_sec"]) for s in states], dtype=np.float64)
    raw_xyz = np.array([s["translation"] for s in states], dtype=np.float64)
    raw_yaw = np.unwrap(np.array([s["yaw"] for s in states], dtype=np.float64))

    positions, headings, _, valid_list = _empty_series(num_steps)
    interp_xy = np.full((num_steps, 2), ERR_VAL, dtype=np.float64)
    valid = np.zeros(num_steps, dtype=bool)

    for out_idx, target_t in enumerate(target_times_sec):
        right_idx = int(np.searchsorted(raw_t, target_t, side="left"))
        if not _is_interpolation_gap_valid(
            raw_t, target_t, right_idx, config.max_annotation_gap_sec
        ):
            continue

        xyz = _interpolate_value(raw_t, raw_xyz, target_t, right_idx)
        yaw = _interpolate_value(raw_t, raw_yaw, target_t, right_idx)
        if xyz is None or yaw is None:
            continue

        positions[out_idx] = {
            "x": float(xyz[0]),
            "y": float(xyz[1]),
            "z": float(xyz[2]),
        }
        headings[out_idx] = float(wrap_yaw(float(yaw)))
        interp_xy[out_idx] = xyz[:2]
        valid[out_idx] = True

    if config.require_start_valid and not valid[0]:
        return None
    if int(valid.sum()) < config.min_valid_steps:
        return None

    valid_indices = np.flatnonzero(valid)
    last_valid_idx = int(valid_indices[-1])
    velocities = _compute_velocities(interp_xy, valid, config.dt)

    valid_list = [bool(v) for v in valid]
    size = np.median(np.array([s["size"] for s in states], dtype=np.float64), axis=0)
    return {
        "position": positions,
        "width": float(size[0]),
        "length": float(size[1]),
        "height": float(size[2]),
        "heading": headings,
        "velocity": velocities,
        "valid": valid_list,
        "goalPosition": positions[last_valid_idx],
    }


def _sample_time_sec(sample: dict[str, Any]) -> float:
    return float(sample["timestamp"]) / 1e6


def collect_scene_samples(nusc: Any, scene: dict[str, Any]) -> list[dict[str, Any]]:
    """Collect all keyframe samples for one NuScenes scene."""

    samples = []
    sample_token = scene["first_sample_token"]
    while sample_token:
        sample = nusc.get("sample", sample_token)
        samples.append(sample)
        sample_token = sample["next"]
    return samples


def collect_annotation_tracks(
    nusc: Any,
    samples: list[dict[str, Any]],
) -> dict[str, list[dict[str, Any]]]:
    """Collect NuScenes sample annotations into per-instance tracks."""

    tracks: dict[str, list[dict[str, Any]]] = defaultdict(list)
    for sample in samples:
        time_sec = _sample_time_sec(sample)
        for ann_token in sample["anns"]:
            ann = nusc.get("sample_annotation", ann_token)
            gpudrive_type = category_to_gpudrive(ann["category_name"])
            if gpudrive_type is None:
                continue

            tracks[ann["instance_token"]].append(
                {
                    "time_sec": time_sec,
                    "translation": ann["translation"],
                    "yaw": yaw_from_quaternion(ann["rotation"]),
                    "size": ann["size"],
                    "category_name": ann["category_name"],
                    "gpudrive_type": gpudrive_type,
                }
            )

    return tracks


def collect_ego_track(
    nusc: Any,
    samples: list[dict[str, Any]],
    config: NuScenesAdapterConfig,
) -> list[dict[str, Any]]:
    """Collect ego poses as a synthetic vehicle track."""

    ego_track = []
    ego_size = [config.ego_width, config.ego_length, config.ego_height]
    for sample in samples:
        lidar_token = sample["data"].get("LIDAR_TOP")
        if lidar_token is None:
            continue

        sample_data = nusc.get("sample_data", lidar_token)
        pose = nusc.get("ego_pose", sample_data["ego_pose_token"])
        ego_track.append(
            {
                "time_sec": _sample_time_sec(sample),
                "translation": pose["translation"],
                "yaw": yaw_from_quaternion(pose["rotation"]),
                "size": ego_size,
                "category_name": "vehicle.ego",
                "gpudrive_type": "vehicle",
            }
        )
    return ego_track


def iter_window_starts(
    samples: list[dict[str, Any]],
    config: NuScenesAdapterConfig,
    window_stride_sec: Optional[float] = None,
) -> list[float]:
    """Return conversion window starts in absolute NuScenes seconds."""

    if not samples:
        return []

    first_time = _sample_time_sec(samples[0])
    last_time = _sample_time_sec(samples[-1])
    latest_start = last_time - config.horizon_sec

    if config.require_full_window and latest_start < first_time:
        return []

    if window_stride_sec is None or window_stride_sec <= 0:
        return [first_time]

    latest_allowed_start = latest_start if config.require_full_window else last_time
    starts = []
    cur = first_time
    while cur <= latest_allowed_start + 1e-6:
        starts.append(cur)
        cur += window_stride_sec
    return starts


def _geometry_point(x: float, y: float, z: float = 0.0) -> dict[str, float]:
    return {"x": float(x), "y": float(y), "z": float(z)}


def _dedupe_points(points: list[dict[str, float]]) -> list[dict[str, float]]:
    if not points:
        return points

    deduped = [points[0]]
    for point in points[1:]:
        prev = deduped[-1]
        if abs(point["x"] - prev["x"]) > 1e-6 or abs(point["y"] - prev["y"]) > 1e-6:
            deduped.append(point)

    if len(deduped) > 2:
        first = deduped[0]
        last = deduped[-1]
        if abs(first["x"] - last["x"]) <= 1e-6 and abs(first["y"] - last["y"]) <= 1e-6:
            deduped.pop()
    return deduped


def _thin_points(
    points: list[dict[str, float]],
    max_points: int,
) -> list[dict[str, float]]:
    if len(points) <= max_points:
        return points

    keep = np.linspace(0, len(points) - 1, max_points).round().astype(int)
    return [points[int(i)] for i in keep]


def _coords_to_geometry(
    coords: Iterable[Any],
    max_points: int,
) -> list[dict[str, float]]:
    points = []
    for coord in coords:
        x = coord[0]
        y = coord[1]
        points.append(_geometry_point(x, y, 0.0))
    return _thin_points(_dedupe_points(points), max_points)


def _add_road(
    roads: list[dict[str, Any]],
    geometry: list[dict[str, float]],
    road_type: str,
    map_element_id: int,
    road_id: int,
    config: NuScenesAdapterConfig,
) -> int:
    if len(roads) >= config.max_road_elements:
        return road_id

    min_points = 4 if road_type in {"crosswalk", "speed_bump"} else 2
    if road_type == "stop_sign":
        min_points = 1
    if len(geometry) < min_points:
        return road_id

    roads.append(
        {
            "geometry": geometry,
            "type": road_type,
            "map_element_id": int(map_element_id),
            "id": road_id,
        }
    )
    return road_id + 1


def _get_records_in_radius(
    nusc_map: Any,
    center_xy: tuple[float, float],
    radius_m: float,
    layer_names: list[str],
) -> dict[str, list[str]]:
    try:
        return nusc_map.get_records_in_radius(
            center_xy[0],
            center_xy[1],
            radius_m,
            layer_names,
            mode="intersect",
        )
    except TypeError:  # pragma: no cover - SDK-version compatibility.
        return nusc_map.get_records_in_radius(
            center_xy[0],
            center_xy[1],
            radius_m,
            layer_names,
        )


def _discretize_lanes(
    nusc_map: Any,
    lane_tokens: list[str],
    discretization_m: float,
) -> dict[str, Any]:
    try:
        return nusc_map.discretize_lanes(
            lane_tokens, resolution_meters=discretization_m
        )
    except TypeError:  # pragma: no cover - SDK-version compatibility.
        return nusc_map.discretize_lanes(
            lane_tokens, discretization_meters=discretization_m
        )


def _extract_polygon_exterior(nusc_map: Any, polygon_token: str) -> list[Any]:
    polygon = nusc_map.extract_polygon(polygon_token)
    return list(polygon.exterior.coords)


def _extract_line_coords(nusc_map: Any, line_token: str) -> list[Any]:
    line = nusc_map.extract_line(line_token)
    return list(line.coords)


def _road_center_from_objects(objects: list[dict[str, Any]]) -> tuple[float, float]:
    starts = []
    for obj in objects:
        for idx, valid in enumerate(obj["valid"]):
            if valid:
                pos = obj["position"][idx]
                starts.append((pos["x"], pos["y"]))
                break

    if not starts:
        return 0.0, 0.0
    arr = np.array(starts, dtype=np.float64)
    return float(arr[:, 0].mean()), float(arr[:, 1].mean())


def load_nuscenes_map(dataroot: str, map_name: str):
    """Load a NuScenes map object."""

    try:
        from nuscenes.map_expansion.map_api import NuScenesMap
    except ImportError as exc:  # pragma: no cover - depends on local install.
        raise ImportError(
            "NuScenes map conversion requires `nuscenes-devkit`."
        ) from exc

    return NuScenesMap(dataroot=dataroot, map_name=map_name)


def get_scene_map_name(nusc: Any, scene: dict[str, Any]) -> str:
    log = nusc.get("log", scene["log_token"])
    return log["location"]


def extract_roads(
    nusc_map: Any,
    center_xy: tuple[float, float],
    config: NuScenesAdapterConfig,
) -> list[dict[str, Any]]:
    """Extract nearby NuScenes map layers as GPUDrive road objects."""

    layer_names = [
        "lane",
        "lane_connector",
        "road_divider",
        "lane_divider",
        "drivable_area",
        "ped_crossing",
        "stop_line",
    ]
    records = _get_records_in_radius(
        nusc_map, center_xy, config.map_radius_m, layer_names
    )

    roads: list[dict[str, Any]] = []
    road_id = 0

    lane_tokens = records.get("lane", []) + records.get("lane_connector", [])
    if lane_tokens:
        lane_paths = _discretize_lanes(
            nusc_map, lane_tokens, config.map_discretization_m
        )
        for points in lane_paths.values():
            geometry = _coords_to_geometry(points, config.max_geometry_points)
            road_id = _add_road(
                roads,
                geometry,
                "lane",
                MapElementIds.LANE_SURFACE_STREET,
                road_id,
                config,
            )

    for layer_name, map_id in [
        ("road_divider", MapElementIds.ROAD_LINE_SOLID_SINGLE_WHITE),
        ("lane_divider", MapElementIds.ROAD_LINE_BROKEN_SINGLE_WHITE),
        ("stop_line", MapElementIds.ROAD_LINE_UNKNOWN),
    ]:
        for token in records.get(layer_name, []):
            record = nusc_map.get(layer_name, token)
            line_token = record.get("line_token")
            if not line_token:
                continue
            geometry = _coords_to_geometry(
                _extract_line_coords(nusc_map, line_token),
                config.max_geometry_points,
            )
            road_id = _add_road(
                roads,
                geometry,
                "road_line",
                map_id,
                road_id,
                config,
            )

    for token in records.get("drivable_area", []):
        record = nusc_map.get("drivable_area", token)
        for polygon_token in record.get("polygon_tokens", []):
            geometry = _coords_to_geometry(
                _extract_polygon_exterior(nusc_map, polygon_token),
                config.max_geometry_points,
            )
            road_id = _add_road(
                roads,
                geometry,
                "road_edge",
                MapElementIds.ROAD_EDGE_BOUNDARY,
                road_id,
                config,
            )

    for token in records.get("ped_crossing", []):
        record = nusc_map.get("ped_crossing", token)
        polygon_token = record.get("polygon_token")
        if not polygon_token:
            continue
        geometry = _coords_to_geometry(
            _extract_polygon_exterior(nusc_map, polygon_token),
            config.max_geometry_points,
        )
        road_id = _add_road(
            roads,
            geometry,
            "crosswalk",
            MapElementIds.CROSSWALK,
            road_id,
            config,
        )

    return roads


def _make_gpudrive_object(
    obj_id: int,
    gpudrive_type: str,
    states: list[dict[str, Any]],
    target_times_sec: np.ndarray,
    config: NuScenesAdapterConfig,
    mark_as_expert: bool = False,
) -> Optional[dict[str, Any]]:
    obj = interpolate_track(states, target_times_sec, config)
    if obj is None:
        return None

    obj["type"] = gpudrive_type
    obj["id"] = obj_id
    obj["mark_as_expert"] = mark_as_expert
    return obj


def scene_to_gpudrive_scenarios(
    nusc: Any,
    scene: dict[str, Any],
    dataroot: str,
    config: NuScenesAdapterConfig,
    file_prefix: str = "tfrecord-nuscenes",
    window_stride_sec: Optional[float] = None,
    map_cache: Optional[dict[str, Any]] = None,
) -> list[dict[str, Any]]:
    """Convert a NuScenes scene into one or more GPUDrive scenario dicts."""

    samples = collect_scene_samples(nusc, scene)
    window_starts = iter_window_starts(samples, config, window_stride_sec)
    if not window_starts:
        return []

    annotation_tracks = collect_annotation_tracks(nusc, samples)
    ego_track = collect_ego_track(nusc, samples, config) if config.include_ego else []
    map_cache = map_cache if map_cache is not None else {}
    scenarios = []

    for window_idx, window_start in enumerate(window_starts):
        target_times = window_start + np.arange(config.num_steps, dtype=np.float64) * config.dt
        objects = []
        next_id = 0

        if config.include_ego:
            ego_obj = _make_gpudrive_object(
                obj_id=next_id,
                gpudrive_type="vehicle",
                states=ego_track,
                target_times_sec=target_times,
                config=config,
            )
            if ego_obj is not None:
                objects.append(ego_obj)
                next_id += 1

        sorted_tracks = sorted(
            annotation_tracks.items(),
            key=lambda item: (item[1][0]["time_sec"], item[0]),
        )
        for _, states in sorted_tracks:
            gpudrive_type = states[0]["gpudrive_type"]
            obj = _make_gpudrive_object(
                obj_id=next_id,
                gpudrive_type=gpudrive_type,
                states=states,
                target_times_sec=target_times,
                config=config,
            )
            if obj is None:
                continue
            objects.append(obj)
            next_id += 1

        if not objects:
            continue

        center_xy = _road_center_from_objects(objects)
        roads: list[dict[str, Any]] = []
        if config.include_map:
            try:
                map_name = get_scene_map_name(nusc, scene)
                if map_name not in map_cache:
                    map_cache[map_name] = load_nuscenes_map(dataroot, map_name)
                roads = extract_roads(map_cache[map_name], center_xy, config)
            except Exception as exc:  # pragma: no cover - SDK/data dependent.
                if config.fail_on_missing_map:
                    raise
                logging.warning(
                    "Skipping map extraction for scene %s window %d: %s",
                    scene.get("name", scene.get("token")),
                    window_idx,
                    exc,
                )

        sdc_track_index = 0 if config.include_ego and objects and objects[0]["id"] == 0 else -1
        tracks_to_predict = [
            {"track_index": idx, "difficulty": 0}
            for idx, obj in enumerate(objects)
            if obj["type"] in {"vehicle", "cyclist", "pedestrian"}
            and not obj.get("mark_as_expert", False)
        ]
        objects_of_interest = [
            obj["id"]
            for obj in objects
            if obj["type"] in {"vehicle", "cyclist"}
            and not obj.get("mark_as_expert", False)
        ]

        output_name = f"{file_prefix}_{scene['name']}_{window_idx:03d}.json"
        scenarios.append(
            {
                "name": output_name,
                "scenario_id": f"{scene['token'][:24]}_{window_idx:03d}",
                "objects": objects,
                "roads": roads,
                "tl_states": {},
                "metadata": {
                    "sdc_track_index": sdc_track_index,
                    "objects_of_interest": objects_of_interest,
                    "tracks_to_predict": tracks_to_predict,
                    "source": {
                        "dataset": "nuscenes",
                        "scene_name": scene["name"],
                        "scene_token": scene["token"],
                        "window_start_sec": float(window_start),
                        "target_hz": config.target_hz,
                        "num_steps": config.num_steps,
                    },
                },
            }
        )

    return scenarios


def select_scenes(
    nusc: Any,
    split: Optional[str] = None,
    scene_names: Optional[set[str]] = None,
    max_scenes: Optional[int] = None,
) -> list[dict[str, Any]]:
    """Select NuScenes scenes by split/name."""

    selected_names = scene_names
    if split:
        try:
            from nuscenes.utils.splits import create_splits_scenes
        except ImportError as exc:  # pragma: no cover - depends on local install.
            raise ImportError(
                "NuScenes split selection requires `nuscenes-devkit`."
            ) from exc

        splits = create_splits_scenes()
        if split not in splits:
            valid = ", ".join(sorted(splits))
            raise ValueError(f"Unknown NuScenes split '{split}'. Valid splits: {valid}")
        selected_names = set(splits[split])

    scenes = [
        scene
        for scene in nusc.scene
        if selected_names is None or scene["name"] in selected_names
    ]
    if max_scenes is not None:
        scenes = scenes[:max_scenes]
    return scenes


def convert_nuscenes_dataset(
    dataroot: str,
    version: str,
    output_dir: str | Path,
    config: NuScenesAdapterConfig,
    split: Optional[str] = None,
    scene_names: Optional[set[str]] = None,
    max_scenes: Optional[int] = None,
    file_prefix: str = "tfrecord-nuscenes",
    window_stride_sec: Optional[float] = None,
    verbose: bool = True,
) -> dict[str, int]:
    """Convert NuScenes scenes and write GPUDrive JSON files."""

    nusc = load_nuscenes(version=version, dataroot=dataroot, verbose=verbose)
    scenes = select_scenes(
        nusc,
        split=split,
        scene_names=scene_names,
        max_scenes=max_scenes,
    )

    output_path = Path(output_dir)
    output_path.mkdir(parents=True, exist_ok=True)
    map_cache: dict[str, Any] = {}
    written = 0
    skipped = 0

    for scene in scenes:
        scenarios = scene_to_gpudrive_scenarios(
            nusc=nusc,
            scene=scene,
            dataroot=dataroot,
            config=config,
            file_prefix=file_prefix,
            window_stride_sec=window_stride_sec,
            map_cache=map_cache,
        )
        if not scenarios:
            skipped += 1
            continue

        for scenario in scenarios:
            with (output_path / scenario["name"]).open("w", encoding="utf-8") as f:
                json.dump(scenario, f)
            written += 1

    return {
        "scenes_selected": len(scenes),
        "scenes_without_outputs": skipped,
        "files_written": written,
    }