Datasets:

Yoonna
/

tartandrive_2.0

"""Gymnasium environment: drive a MuJoCo ATV over terrain reconstructed from a
single TartanDrive 2.0 sequence.
Observation/action layout is chosen to match the real dataset's own column
conventions (metadata_utils.lidar_terrain's ENU frame, gps_odom's 13-column
layout, IMU's angular-velocity+linear-acceleration 6-column layout) so this
env's rollouts stay directly comparable to real logs, and so a policy trained
here can later be evaluated/fine-tuned against the real `kitti_format` data.
"""
import os
os.environ.setdefault("MUJOCO_GL", "egl")

from pathlib import Path

import gymnasium as gym
import mujoco
import numpy as np
from gymnasium import spaces

from metadata_utils.calibration import load_cam_P, load_extrinsics
from metadata_utils.lidar_terrain import build_terrain_for_sequence, load_sequence_poses
from sim.mjcf_builder import CAM_NAME, build_mjcf
from sim.vehicle_config import get_vehicle_config

REPO_ROOT = Path(__file__).resolve().parent.parent
DEFAULT_CALIBRATION_DIR = REPO_ROOT / "calibration"


class TartanDriveEnv(gym.Env):
    """One episode = one drive over the reconstructed terrain patch for `scene_dir`."""

    metadata = {"render_modes": ["rgb_array"]}

    def __init__(self, scene_dir, calibration_dir=None, cache_path=None,
                 frame_stride=5, cell_size=0.3, frame_range=None,
                 control_substeps=4, max_episode_steps=1000,
                 max_tilt_rad=1.0, image_width=1024, image_height=544):
        super().__init__()
        self.scene_dir = Path(scene_dir)
        calibration_dir = Path(calibration_dir or DEFAULT_CALIBRATION_DIR)
        self.extrinsics = load_extrinsics(calibration_dir / "extrinsics.yaml")
        self.cam_P = load_cam_P(calibration_dir / "multisense_intrinsics.txt")
        self.vehicle_cfg = get_vehicle_config()

        self.terrain = build_terrain_for_sequence(
            self.scene_dir, self.extrinsics, cache_path=cache_path,
            frame_stride=frame_stride, cell_size=cell_size, frame_range=frame_range)
        self.poses = load_sequence_poses(self.scene_dir)

        scene = build_mjcf(
            self.terrain, self.vehicle_cfg, self.extrinsics, self.cam_P,
            image_height=image_height, image_width=image_width,
            spawn_xy=(self.poses.positions[0, 0], self.poses.positions[0, 1]),
            spawn_quat_xyzw=self.poses.quats_xyzw[0])
        self.model = mujoco.MjModel.from_xml_string(scene.xml)
        self.model.hfield_data[:] = scene.hfield_data_flat
        self.data = mujoco.MjData(self.model)
        self.renderer = mujoco.Renderer(self.model, height=image_height, width=image_width)

        self._chassis_id = self.model.body("chassis").id
        self._free_qpos_adr = self.model.jnt_qposadr[self.model.joint("chassis_free").id]
        self._free_dof_adr = self.model.jnt_dofadr[self.model.joint("chassis_free").id]

        self.control_substeps = control_substeps
        self.max_episode_steps = max_episode_steps
        self.max_tilt_rad = max_tilt_rad
        self._step_count = 0

        self.observation_space = spaces.Dict({
            "image": spaces.Box(0, 255, (image_height, image_width, 3), dtype=np.uint8),
            "imu": spaces.Box(-np.inf, np.inf, (6,), dtype=np.float32),
            "odom": spaces.Box(-np.inf, np.inf, (13,), dtype=np.float32),
        })
        self.action_space = spaces.Box(
            low=np.array([-1.0, -1.0], dtype=np.float32),
            high=np.array([1.0, 1.0], dtype=np.float32))

    def reset(self, seed=None, options=None):
        super().reset(seed=seed)
        mujoco.mj_resetData(self.model, self.data)
        mujoco.mj_forward(self.model, self.data)
        self.data.ctrl[:] = 0.0
        # settle onto the terrain under gravity before the episode starts, so the
        # initial spawn-height estimate's contact penetration doesn't resolve into
        # a violent first-step impulse on real (bumpy) reconstructed terrain
        for _ in range(150):
            mujoco.mj_step(self.model, self.data)
        self._step_count = 0
        return self._get_obs(), {}

    def step(self, action):
        action = np.clip(np.asarray(action, dtype=np.float64), -1.0, 1.0)
        throttle, steer = action
        self.data.ctrl[0:4] = throttle
        self.data.ctrl[4:6] = steer * self.vehicle_cfg.max_steer_rad

        for _ in range(self.control_substeps):
            mujoco.mj_step(self.model, self.data)
        self._step_count += 1

        obs = self._get_obs()
        terminated = self._check_terminated()
        truncated = self._step_count >= self.max_episode_steps
        reward = 0.0  # stubbed; traversability-based reward is Phase 2 scope
        return obs, reward, terminated, truncated, {}

    def render(self):
        self.renderer.update_scene(self.data, camera=CAM_NAME)
        return self.renderer.render()

    def close(self):
        self.renderer.close()

    # -- internals ------------------------------------------------------------

    def _get_obs(self):
        image = self.render()

        gyro = self.data.sensor("imu_gyro").data
        accel = self.data.sensor("imu_accel").data
        imu = np.concatenate([gyro, accel]).astype(np.float32)

        pos = self.data.qpos[self._free_qpos_adr:self._free_qpos_adr + 3]
        quat_wxyz = self.data.qpos[self._free_qpos_adr + 3:self._free_qpos_adr + 7]
        quat_xyzw = quat_wxyz[[1, 2, 3, 0]]
        linvel_world = self.data.qvel[self._free_dof_adr:self._free_dof_adr + 3]
        angvel_body = self.data.qvel[self._free_dof_adr + 3:self._free_dof_adr + 6]

        # gps_odom convention: col0=northing(+y), col1=-easting(-x), col2=altitude(z)
        odom = np.concatenate([
            [pos[1], -pos[0], pos[2]],
            quat_xyzw, linvel_world, angvel_body,
        ]).astype(np.float32)

        return {"image": image, "imu": imu, "odom": odom}

    def _check_terminated(self):
        R = self.data.xmat[self._chassis_id].reshape(3, 3)
        up_z = R[2, 2]  # world-z component of the chassis' local up axis
        tilt = np.arccos(np.clip(up_z, -1.0, 1.0))
        return bool(tilt > self.max_tilt_rad)

	"""Gymnasium environment: drive a MuJoCo ATV over terrain reconstructed from a
	single TartanDrive 2.0 sequence.

	Observation/action layout is chosen to match the real dataset's own column
	conventions (metadata_utils.lidar_terrain's ENU frame, gps_odom's 13-column
	layout, IMU's angular-velocity+linear-acceleration 6-column layout) so this
	env's rollouts stay directly comparable to real logs, and so a policy trained
	here can later be evaluated/fine-tuned against the real `kitti_format` data.
	"""
	import os
	os.environ.setdefault("MUJOCO_GL", "egl")

	from pathlib import Path

	import gymnasium as gym
	import mujoco
	import numpy as np
	from gymnasium import spaces

	from metadata_utils.calibration import load_cam_P, load_extrinsics
	from metadata_utils.lidar_terrain import build_terrain_for_sequence, load_sequence_poses
	from sim.mjcf_builder import CAM_NAME, build_mjcf
	from sim.vehicle_config import get_vehicle_config

	REPO_ROOT = Path(__file__).resolve().parent.parent
	DEFAULT_CALIBRATION_DIR = REPO_ROOT / "calibration"


	class TartanDriveEnv(gym.Env):
	"""One episode = one drive over the reconstructed terrain patch for `scene_dir`."""

	metadata = {"render_modes": ["rgb_array"]}

	def __init__(self, scene_dir, calibration_dir=None, cache_path=None,
	frame_stride=5, cell_size=0.3, frame_range=None,
	control_substeps=4, max_episode_steps=1000,
	max_tilt_rad=1.0, image_width=1024, image_height=544):
	super().__init__()
	self.scene_dir = Path(scene_dir)
	calibration_dir = Path(calibration_dir or DEFAULT_CALIBRATION_DIR)
	self.extrinsics = load_extrinsics(calibration_dir / "extrinsics.yaml")
	self.cam_P = load_cam_P(calibration_dir / "multisense_intrinsics.txt")
	self.vehicle_cfg = get_vehicle_config()

	self.terrain = build_terrain_for_sequence(
	self.scene_dir, self.extrinsics, cache_path=cache_path,
	frame_stride=frame_stride, cell_size=cell_size, frame_range=frame_range)
	self.poses = load_sequence_poses(self.scene_dir)

	scene = build_mjcf(
	self.terrain, self.vehicle_cfg, self.extrinsics, self.cam_P,
	image_height=image_height, image_width=image_width,
	spawn_xy=(self.poses.positions[0, 0], self.poses.positions[0, 1]),
	spawn_quat_xyzw=self.poses.quats_xyzw[0])
	self.model = mujoco.MjModel.from_xml_string(scene.xml)
	self.model.hfield_data[:] = scene.hfield_data_flat
	self.data = mujoco.MjData(self.model)
	self.renderer = mujoco.Renderer(self.model, height=image_height, width=image_width)

	self._chassis_id = self.model.body("chassis").id
	self._free_qpos_adr = self.model.jnt_qposadr[self.model.joint("chassis_free").id]
	self._free_dof_adr = self.model.jnt_dofadr[self.model.joint("chassis_free").id]

	self.control_substeps = control_substeps
	self.max_episode_steps = max_episode_steps
	self.max_tilt_rad = max_tilt_rad
	self._step_count = 0

	self.observation_space = spaces.Dict({
	"image": spaces.Box(0, 255, (image_height, image_width, 3), dtype=np.uint8),
	"imu": spaces.Box(-np.inf, np.inf, (6,), dtype=np.float32),
	"odom": spaces.Box(-np.inf, np.inf, (13,), dtype=np.float32),
	})
	self.action_space = spaces.Box(
	low=np.array([-1.0, -1.0], dtype=np.float32),
	high=np.array([1.0, 1.0], dtype=np.float32))

	def reset(self, seed=None, options=None):
	super().reset(seed=seed)
	mujoco.mj_resetData(self.model, self.data)
	mujoco.mj_forward(self.model, self.data)
	self.data.ctrl[:] = 0.0
	# settle onto the terrain under gravity before the episode starts, so the
	# initial spawn-height estimate's contact penetration doesn't resolve into
	# a violent first-step impulse on real (bumpy) reconstructed terrain
	for _ in range(150):
	mujoco.mj_step(self.model, self.data)
	self._step_count = 0
	return self._get_obs(), {}

	def step(self, action):
	action = np.clip(np.asarray(action, dtype=np.float64), -1.0, 1.0)
	throttle, steer = action
	self.data.ctrl[0:4] = throttle
	self.data.ctrl[4:6] = steer * self.vehicle_cfg.max_steer_rad

	for _ in range(self.control_substeps):
	mujoco.mj_step(self.model, self.data)
	self._step_count += 1

	obs = self._get_obs()
	terminated = self._check_terminated()
	truncated = self._step_count >= self.max_episode_steps
	reward = 0.0 # stubbed; traversability-based reward is Phase 2 scope
	return obs, reward, terminated, truncated, {}

	def render(self):
	self.renderer.update_scene(self.data, camera=CAM_NAME)
	return self.renderer.render()

	def close(self):
	self.renderer.close()

	# -- internals ------------------------------------------------------------

	def _get_obs(self):
	image = self.render()

	gyro = self.data.sensor("imu_gyro").data
	accel = self.data.sensor("imu_accel").data
	imu = np.concatenate([gyro, accel]).astype(np.float32)

	pos = self.data.qpos[self._free_qpos_adr:self._free_qpos_adr + 3]
	quat_wxyz = self.data.qpos[self._free_qpos_adr + 3:self._free_qpos_adr + 7]
	quat_xyzw = quat_wxyz[[1, 2, 3, 0]]
	linvel_world = self.data.qvel[self._free_dof_adr:self._free_dof_adr + 3]
	angvel_body = self.data.qvel[self._free_dof_adr + 3:self._free_dof_adr + 6]

	# gps_odom convention: col0=northing(+y), col1=-easting(-x), col2=altitude(z)
	odom = np.concatenate([
	[pos[1], -pos[0], pos[2]],
	quat_xyzw, linvel_world, angvel_body,
	]).astype(np.float32)

	return {"image": image, "imu": imu, "odom": odom}

	def _check_terminated(self):
	R = self.data.xmat[self._chassis_id].reshape(3, 3)
	up_z = R[2, 2] # world-z component of the chassis' local up axis
	tilt = np.arccos(np.clip(up_z, -1.0, 1.0))
	return bool(tilt > self.max_tilt_rad)