#!/usr/bin/env python3 """Verify the Phase-1 TartanDrive MuJoCo environment end-to-end. Builds a TartanDriveEnv for one sequence, runs a scripted rollout, saves a few rendered camera frames plus a top-down overlay of the sim trajectory vs. the real GPS track (on the Gascola orthophoto), and optionally runs a fast reset/step smoke-test. Success is qualitative for this PoC: the env should reset/step without exceptions, observations must match observation_space, and the vehicle should drive a plausible distance without tunneling through terrain or diverging (NaNs) -- it is not expected to numerically retrace the real GPS track. """ import argparse import os import subprocess import sys os.environ.setdefault("MUJOCO_GL", "egl") from pathlib import Path REPO_ROOT = Path(__file__).resolve().parent.parent if str(REPO_ROOT) not in sys.path: sys.path.insert(0, str(REPO_ROOT)) import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import numpy as np from PIL import Image from metadata_utils.lidar_terrain import load_sequence_poses from sim.env import TartanDriveEnv def try_pull_atv_scan(): scan_path = REPO_ROOT / "assets" / "vicky1_clean.pts" if scan_path.exists(): return print("[verify_sim_env] assets/vicky1_clean.pts not found; attempting download " "(falls back to generic UTV dimensions in sim/vehicle_config.py if unreachable)...") try: subprocess.run([sys.executable, str(REPO_ROOT / "scripts" / "pull_atv_scan.py")], cwd=REPO_ROOT, timeout=20, capture_output=True, check=True) print("[verify_sim_env] fetched vicky1_clean.pts") except Exception as e: print(f"[verify_sim_env] could not fetch vicky1_clean.pts ({e}); " f"using generic UTV approximation") def run_smoke_test(env): obs, info = env.reset() assert env.observation_space.contains(obs), "reset() obs violates observation_space" assert env.action_space.shape == (2,) for _ in range(20): action = env.action_space.sample() obs, reward, terminated, truncated, info = env.step(action) assert env.observation_space.contains(obs), "step() obs violates observation_space" assert np.isfinite(obs["odom"]).all(), "odom contains NaN/inf" assert np.isfinite(obs["imu"]).all(), "imu contains NaN/inf" if terminated or truncated: obs, info = env.reset() print("[smoke-test] OK: reset/step observations match observation_space, no NaNs") def rollout(env, steps, frame_every=50): obs, info = env.reset() xy = [env.data.qpos[env._free_qpos_adr:env._free_qpos_adr + 2].copy()] frames = [(0, obs["image"].copy())] ended_at = None for i in range(1, steps + 1): throttle, steer = 0.35, 0.25 * np.sin(i / 60.0) obs, reward, terminated, truncated, info = env.step( np.array([throttle, steer], dtype=np.float32)) xy.append(env.data.qpos[env._free_qpos_adr:env._free_qpos_adr + 2].copy()) if i % frame_every == 0: frames.append((i, obs["image"].copy())) if terminated or truncated: ended_at = i print(f"[rollout] episode ended at step {i} " f"(terminated={terminated}, truncated={truncated})") break return np.array(xy), frames, ended_at def save_frames(frames, out_dir): out_dir.mkdir(parents=True, exist_ok=True) for i, img in frames: Image.fromarray(img).save(out_dir / f"frame_{i:05d}.png") def save_overlay(env, sim_xy, out_path): poses = load_sequence_poses(env.scene_dir) real_xy = poses.positions[:, :2] # (easting, northing), relative to sequence start fig, ax = plt.subplots(figsize=(8, 8)) try: import rasterio with rasterio.open(REPO_ROOT / "assets" / "gascola.tif") as ds: img = ds.read([1, 2, 3]).transpose(1, 2, 0) left, top = ds.transform * (0, 0) right, bottom = ds.transform * (ds.width, ds.height) utm0 = env.terrain.utm_origin extent = [left - utm0[0], right - utm0[0], bottom - utm0[1], top - utm0[1]] ax.imshow(img, extent=extent) except Exception as e: print(f"[overlay] could not load gascola.tif basemap ({e}); plotting trajectories only") ax.plot(real_xy[:, 0], real_xy[:, 1], "-", color="0.6", lw=1.0, label="real GPS track (full)") ax.plot(sim_xy[:, 0], sim_xy[:, 1], "-", color="tab:red", lw=2.0, label="sim rollout") ax.plot(0, 0, "k*", ms=14, label="sequence start") ax.set_xlabel("relative easting [m]") ax.set_ylabel("relative northing [m]") ax.legend(loc="best") ax.set_aspect("equal") margin = 15 ax.set_xlim(sim_xy[:, 0].min() - margin, sim_xy[:, 0].max() + margin) ax.set_ylim(sim_xy[:, 1].min() - margin, sim_xy[:, 1].max() + margin) fig.tight_layout() fig.savefig(out_path, dpi=150) plt.close(fig) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--sequence", default="2023-10-26-14-42-35_turnpike_afternoon_fall") parser.add_argument("--kitti-root", default=str(REPO_ROOT / "kitti_format")) parser.add_argument("--steps", type=int, default=350) parser.add_argument("--frame-stride", type=int, default=3, help="use every Nth lidar frame when aggregating terrain") parser.add_argument("--frame-range-end", type=int, default=300, help="restrict terrain aggregation to frames [0, this) for a fast PoC build") parser.add_argument("--cell-size", type=float, default=0.3) parser.add_argument("--out", default=str(REPO_ROOT / "sim_verify_out")) parser.add_argument("--smoke-test", action="store_true") args = parser.parse_args() try_pull_atv_scan() scene_dir = Path(args.kitti_root) / args.sequence out_dir = Path(args.out) out_dir.mkdir(parents=True, exist_ok=True) cache_path = out_dir / f"{args.sequence}_terrain.npz" env = TartanDriveEnv( scene_dir=scene_dir, cache_path=cache_path, frame_stride=args.frame_stride, cell_size=args.cell_size, frame_range=(0, args.frame_range_end), max_episode_steps=args.steps + 1) if args.smoke_test: run_smoke_test(env) sim_xy, frames, ended_at = rollout(env, args.steps) save_frames(frames, out_dir / "frames") save_overlay(env, sim_xy, out_dir / "trajectory_overlay.png") env.close() distance = float(np.linalg.norm(sim_xy[-1] - sim_xy[0])) status = "SUCCESS" if ended_at is None else "ENDED EARLY (tilt/step limit)" print(f"[verify_sim_env] {status}: {len(sim_xy) - 1} steps, " f"net displacement {distance:.2f} m, {len(frames)} frames saved to {out_dir}") if __name__ == "__main__": main()