flip_landing_reward_wrapper.py
| 1 | |
| 2 | from __future__ import annotations |
| 3 | |
| 4 | from collections.abc import Mapping |
| 5 | import math |
| 6 | |
| 7 | import gymnasium as gym |
| 8 | import numpy as np |
| 9 | |
| 10 | |
| 11 | |
| 12 | DEFAULT_REWARD_CONFIG = { |
| 13 | # ------------------------------------------------------------------ |
| 14 | # 1. Task-stage definitions |
| 15 | # |
| 16 | # These values define when the wrapper considers the flip and |
| 17 | # post-flip recovery stages complete. |
| 18 | # ------------------------------------------------------------------ |
| 19 | |
| 20 | "required_rotations": 1.0, |
| 21 | "rotation_direction": 1, |
| 22 | "upright_tolerance_radians": 0.30, |
| 23 | "recovery_angular_velocity_tolerance": 0.50, |
| 24 | |
| 25 | # ------------------------------------------------------------------ |
| 26 | # 2. Original LunarLander reward contribution |
| 27 | # |
| 28 | # The custom reward is added to a weighted version of Gymnasium's |
| 29 | # original LunarLander reward. |
| 30 | # ------------------------------------------------------------------ |
| 31 | |
| 32 | "pre_flip_original_reward_weight": 0.25, |
| 33 | "post_flip_original_reward_weight": 1.0, |
| 34 | |
| 35 | # ------------------------------------------------------------------ |
| 36 | # 2a. Optional pre-flip horizontal shaping |
| 37 | # |
| 38 | # These are zero by default so the existing Phase A-D |
| 39 | # configurations retain their previous behaviour. |
| 40 | # ------------------------------------------------------------------ |
| 41 | |
| 42 | "pre_flip_center_weight": 0.0, |
| 43 | "pre_flip_horizontal_speed_weight": 0.0, |
| 44 | "pre_flip_shaping_weight": 0.0, |
| 45 | "pre_flip_shaping_clip": 10.0, |
| 46 | |
| 47 | # Penalise completing the flip outside a generous |
| 48 | # horizontal corridor around the landing pad. |
| 49 | "flip_corridor_half_width": 0.35, |
| 50 | "flip_entry_penalty_weight": 0.0, |
| 51 | |
| 52 | # ------------------------------------------------------------------ |
| 53 | # 3. Positive stage and terminal rewards |
| 54 | # |
| 55 | # Progress is paid only when the maximum achieved rotation progress |
| 56 | # increases. The other rewards are one-off stage rewards. |
| 57 | # ------------------------------------------------------------------ |
| 58 | |
| 59 | "rotation_progress_bonus": 150.0, |
| 60 | "flip_completion_bonus": 200.0, |
| 61 | "recovery_bonus": 100.0, |
| 62 | "flip_landing_bonus": 750.0, |
| 63 | |
| 64 | # ------------------------------------------------------------------ |
| 65 | # 4. Terminal failure penalties |
| 66 | # ------------------------------------------------------------------ |
| 67 | |
| 68 | "landing_without_flip_penalty": 300.0, |
| 69 | "no_flip_terminal_penalty": 300.0, |
| 70 | "failed_landing_penalty": 150.0, |
| 71 | "outside_zone_landing_penalty": 1_000.0, |
| 72 | "in_zone_crash_penalty": 1_200.0, |
| 73 | |
| 74 | # ------------------------------------------------------------------ |
| 75 | # 5. Potential-difference shaping controls |
| 76 | # |
| 77 | # These control how changes in post-flip state quality are converted |
| 78 | # into a dense per-step reward. |
| 79 | # ------------------------------------------------------------------ |
| 80 | |
| 81 | "post_flip_shaping_weight": 1.0, |
| 82 | "post_flip_shaping_gamma": 0.995, |
| 83 | "post_flip_shaping_clip": 10.0, |
| 84 | |
| 85 | # ------------------------------------------------------------------ |
| 86 | # 6. Post-flip state-quality weights |
| 87 | # |
| 88 | # These determine the relative importance of position, velocity, |
| 89 | # angular velocity and leg contact in the potential. |
| 90 | # ------------------------------------------------------------------ |
| 91 | |
| 92 | "post_flip_center_weight": 50.0, |
| 93 | "post_flip_horizontal_speed_weight": 30.0, |
| 94 | "post_flip_vertical_speed_weight": 30.0, |
| 95 | "post_flip_angle_weight": 20.0, |
| 96 | "post_flip_angular_speed_weight": 10.0, |
| 97 | "post_flip_leg_contact_weight": 10.0, |
| 98 | |
| 99 | # ------------------------------------------------------------------ |
| 100 | # 7. Horizontal landing-zone guidance |
| 101 | # ------------------------------------------------------------------ |
| 102 | |
| 103 | "landing_zone_half_width": 0.20, # Defines accepted zone (-0.2<=x<= 0.2) |
| 104 | "post_flip_zone_excess_weight": 200.0, # Penalises distance beyond the zone boundaries |
| 105 | "post_flip_target_vx_gain": 0.50, # Converts horizontal position into desired velocity towards the centre |
| 106 | "post_flip_max_target_vx": 0.35, # Caps magnitude of desired horizontal speed |
| 107 | "post_flip_horizontal_deadband": 0.08, # Stops rewarding tiny centering corrections close to the centre |
| 108 | |
| 109 | # ------------------------------------------------------------------ |
| 110 | # 8. Vertical descent and soft-touchdown guidance |
| 111 | # ------------------------------------------------------------------ |
| 112 | |
| 113 | "post_flip_target_vy_high": -0.45, # Desired vertical velocity while sufficiently high |
| 114 | "post_flip_target_vy_near_ground": -0.12, # Desired slower descent near the ground |
| 115 | "near_ground_height": 0.60, # Height below which the overspeed penalty progressively strengthens |
| 116 | "safe_touchdown_vertical_speed": 0.18, # Downward-speed threshold before the quadratic overspeed penalty starts |
| 117 | "near_ground_overspeed_weight": 120.0, # Strength of the quadratic near-ground overspeed penalty |
| 118 | } |
| 119 | |
| 120 | |
| 121 | class FlipLandingRewardWrapper(gym.Wrapper): |
| 122 | """ |
| 123 | Stage-aware LunarLander objective: |
| 124 | |
| 125 | 1. Complete one full rotation in the selected direction. |
| 126 | 2. Arrest the spin and recover to an upright attitude. |
| 127 | 3. Re-centre, stabilise and land safely. |
| 128 | |
| 129 | The original eight-value observation is extended with: |
| 130 | - signed rotation progress in [-1, 1]; |
| 131 | - a flip-completed flag; |
| 132 | - a post-flip-recovery-completed flag. |
| 133 | |
| 134 | Post-flip recovery uses a potential-difference reward based on |
| 135 | horizontal position, horizontal speed, attitude, angular speed, |
| 136 | and leg contact. This rewards improvement rather than repeatedly |
| 137 | paying the agent simply for occupying a favourable state. |
| 138 | """ |
| 139 | |
| 140 | def __init__( |
| 141 | self, |
| 142 | env: gym.Env, |
| 143 | *, |
| 144 | reward_config: ( |
| 145 | Mapping[str, float | int] |
| 146 | | None |
| 147 | ) = None, |
| 148 | ): |
| 149 | """ |
| 150 | Initialise the reward wrapper. |
| 151 | |
| 152 | DEFAULT_REWARD_CONFIG is the single source of default values. |
| 153 | `reward_config` may override any subset of those values for a |
| 154 | particular curriculum phase. |
| 155 | """ |
| 156 | |
| 157 | super().__init__(env) |
| 158 | |
| 159 | supplied_config = dict( |
| 160 | reward_config or {} |
| 161 | ) |
| 162 | |
| 163 | # Reject misspelled or obsolete parameter names. Without this |
| 164 | # check, an accidental typo could silently leave the default |
| 165 | # value active. |
| 166 | unknown_parameters = ( |
| 167 | set(supplied_config) |
| 168 | - set(DEFAULT_REWARD_CONFIG) |
| 169 | ) |
| 170 | |
| 171 | if unknown_parameters: |
| 172 | raise ValueError( |
| 173 | "Unknown reward parameters: " |
| 174 | f"{sorted(unknown_parameters)}" |
| 175 | ) |
| 176 | |
| 177 | # Make a new dictionary so the global defaults are never mutated. |
| 178 | config = { |
| 179 | **DEFAULT_REWARD_CONFIG, |
| 180 | **supplied_config, |
| 181 | } |
| 182 | |
| 183 | # ============================================================== |
| 184 | # 1. Task-stage definitions |
| 185 | # ============================================================== |
| 186 | |
| 187 | self.required_rotations = float( |
| 188 | config["required_rotations"] |
| 189 | ) |
| 190 | |
| 191 | self.rotation_direction = int( |
| 192 | config["rotation_direction"] |
| 193 | ) |
| 194 | |
| 195 | self.upright_tolerance_radians = float( |
| 196 | config[ |
| 197 | "upright_tolerance_radians" |
| 198 | ] |
| 199 | ) |
| 200 | |
| 201 | self.recovery_angular_velocity_tolerance = float( |
| 202 | config[ |
| 203 | "recovery_angular_velocity_tolerance" |
| 204 | ] |
| 205 | ) |
| 206 | |
| 207 | # ============================================================== |
| 208 | # 2. Original LunarLander reward contribution |
| 209 | # ============================================================== |
| 210 | |
| 211 | self.pre_flip_original_reward_weight = float( |
| 212 | config[ |
| 213 | "pre_flip_original_reward_weight" |
| 214 | ] |
| 215 | ) |
| 216 | |
| 217 | self.post_flip_original_reward_weight = float( |
| 218 | config[ |
| 219 | "post_flip_original_reward_weight" |
| 220 | ] |
| 221 | ) |
| 222 | |
| 223 | # ============================================================== |
| 224 | # 2a. Pre-flip horizontal guidance |
| 225 | # ============================================================== |
| 226 | |
| 227 | self.pre_flip_center_weight = float( |
| 228 | config["pre_flip_center_weight"] |
| 229 | ) |
| 230 | |
| 231 | self.pre_flip_horizontal_speed_weight = float( |
| 232 | config[ |
| 233 | "pre_flip_horizontal_speed_weight" |
| 234 | ] |
| 235 | ) |
| 236 | |
| 237 | self.pre_flip_shaping_weight = float( |
| 238 | config["pre_flip_shaping_weight"] |
| 239 | ) |
| 240 | |
| 241 | self.pre_flip_shaping_clip = float( |
| 242 | config["pre_flip_shaping_clip"] |
| 243 | ) |
| 244 | |
| 245 | self.flip_corridor_half_width = float( |
| 246 | config["flip_corridor_half_width"] |
| 247 | ) |
| 248 | |
| 249 | self.flip_entry_penalty_weight = float( |
| 250 | config["flip_entry_penalty_weight"] |
| 251 | ) |
| 252 | |
| 253 | # ============================================================== |
| 254 | # 3. Positive stage and terminal rewards |
| 255 | # ============================================================== |
| 256 | |
| 257 | self.rotation_progress_bonus = float( |
| 258 | config[ |
| 259 | "rotation_progress_bonus" |
| 260 | ] |
| 261 | ) |
| 262 | |
| 263 | self.flip_completion_bonus = float( |
| 264 | config[ |
| 265 | "flip_completion_bonus" |
| 266 | ] |
| 267 | ) |
| 268 | |
| 269 | self.recovery_bonus = float( |
| 270 | config["recovery_bonus"] |
| 271 | ) |
| 272 | |
| 273 | self.flip_landing_bonus = float( |
| 274 | config["flip_landing_bonus"] |
| 275 | ) |
| 276 | |
| 277 | # ============================================================== |
| 278 | # 4. Terminal failure penalties |
| 279 | # ============================================================== |
| 280 | |
| 281 | self.landing_without_flip_penalty = float( |
| 282 | config[ |
| 283 | "landing_without_flip_penalty" |
| 284 | ] |
| 285 | ) |
| 286 | |
| 287 | self.no_flip_terminal_penalty = float( |
| 288 | config[ |
| 289 | "no_flip_terminal_penalty" |
| 290 | ] |
| 291 | ) |
| 292 | |
| 293 | self.failed_landing_penalty = float( |
| 294 | config[ |
| 295 | "failed_landing_penalty" |
| 296 | ] |
| 297 | ) |
| 298 | |
| 299 | self.outside_zone_landing_penalty = float( |
| 300 | config[ |
| 301 | "outside_zone_landing_penalty" |
| 302 | ] |
| 303 | ) |
| 304 | |
| 305 | self.in_zone_crash_penalty = float( |
| 306 | config[ |
| 307 | "in_zone_crash_penalty" |
| 308 | ] |
| 309 | ) |
| 310 | |
| 311 | # ============================================================== |
| 312 | # 5. Potential-difference shaping controls |
| 313 | # ============================================================== |
| 314 | |
| 315 | self.post_flip_shaping_weight = float( |
| 316 | config[ |
| 317 | "post_flip_shaping_weight" |
| 318 | ] |
| 319 | ) |
| 320 | |
| 321 | self.post_flip_shaping_gamma = float( |
| 322 | config[ |
| 323 | "post_flip_shaping_gamma" |
| 324 | ] |
| 325 | ) |
| 326 | |
| 327 | self.post_flip_shaping_clip = float( |
| 328 | config[ |
| 329 | "post_flip_shaping_clip" |
| 330 | ] |
| 331 | ) |
| 332 | |
| 333 | # ============================================================== |
| 334 | # 6. Post-flip state-quality weights |
| 335 | # ============================================================== |
| 336 | |
| 337 | self.post_flip_center_weight = float( |
| 338 | config[ |
| 339 | "post_flip_center_weight" |
| 340 | ] |
| 341 | ) |
| 342 | |
| 343 | self.post_flip_horizontal_speed_weight = float( |
| 344 | config[ |
| 345 | "post_flip_horizontal_speed_weight" |
| 346 | ] |
| 347 | ) |
| 348 | |
| 349 | self.post_flip_vertical_speed_weight = float( |
| 350 | config[ |
| 351 | "post_flip_vertical_speed_weight" |
| 352 | ] |
| 353 | ) |
| 354 | |
| 355 | self.post_flip_angle_weight = float( |
| 356 | config[ |
| 357 | "post_flip_angle_weight" |
| 358 | ] |
| 359 | ) |
| 360 | |
| 361 | self.post_flip_angular_speed_weight = float( |
| 362 | config[ |
| 363 | "post_flip_angular_speed_weight" |
| 364 | ] |
| 365 | ) |
| 366 | |
| 367 | self.post_flip_leg_contact_weight = float( |
| 368 | config[ |
| 369 | "post_flip_leg_contact_weight" |
| 370 | ] |
| 371 | ) |
| 372 | |
| 373 | # ============================================================== |
| 374 | # 7. Horizontal landing-zone guidance |
| 375 | # ============================================================== |
| 376 | |
| 377 | self.landing_zone_half_width = float( |
| 378 | config[ |
| 379 | "landing_zone_half_width" |
| 380 | ] |
| 381 | ) |
| 382 | |
| 383 | self.post_flip_zone_excess_weight = float( |
| 384 | config[ |
| 385 | "post_flip_zone_excess_weight" |
| 386 | ] |
| 387 | ) |
| 388 | |
| 389 | self.post_flip_target_vx_gain = float( |
| 390 | config[ |
| 391 | "post_flip_target_vx_gain" |
| 392 | ] |
| 393 | ) |
| 394 | |
| 395 | self.post_flip_max_target_vx = float( |
| 396 | config[ |
| 397 | "post_flip_max_target_vx" |
| 398 | ] |
| 399 | ) |
| 400 | |
| 401 | self.post_flip_horizontal_deadband = float( |
| 402 | config[ |
| 403 | "post_flip_horizontal_deadband" |
| 404 | ] |
| 405 | ) |
| 406 | |
| 407 | # ============================================================== |
| 408 | # 8. Vertical descent and soft-touchdown guidance |
| 409 | # ============================================================== |
| 410 | |
| 411 | self.post_flip_target_vy_high = float( |
| 412 | config[ |
| 413 | "post_flip_target_vy_high" |
| 414 | ] |
| 415 | ) |
| 416 | |
| 417 | self.post_flip_target_vy_near_ground = float( |
| 418 | config[ |
| 419 | "post_flip_target_vy_near_ground" |
| 420 | ] |
| 421 | ) |
| 422 | |
| 423 | self.near_ground_height = float( |
| 424 | config[ |
| 425 | "near_ground_height" |
| 426 | ] |
| 427 | ) |
| 428 | |
| 429 | self.safe_touchdown_vertical_speed = float( |
| 430 | config[ |
| 431 | "safe_touchdown_vertical_speed" |
| 432 | ] |
| 433 | ) |
| 434 | |
| 435 | self.near_ground_overspeed_weight = float( |
| 436 | config[ |
| 437 | "near_ground_overspeed_weight" |
| 438 | ] |
| 439 | ) |
| 440 | |
| 441 | # ============================================================== |
| 442 | # Configuration validation |
| 443 | # ============================================================== |
| 444 | |
| 445 | if self.required_rotations <= 0.0: |
| 446 | raise ValueError( |
| 447 | "required_rotations must be positive." |
| 448 | ) |
| 449 | |
| 450 | if self.rotation_direction not in ( |
| 451 | -1, |
| 452 | 1, |
| 453 | ): |
| 454 | raise ValueError( |
| 455 | "rotation_direction must be " |
| 456 | "either -1 or 1." |
| 457 | ) |
| 458 | |
| 459 | if not ( |
| 460 | 0.0 |
| 461 | < self.upright_tolerance_radians |
| 462 | <= math.pi |
| 463 | ): |
| 464 | raise ValueError( |
| 465 | "upright_tolerance_radians must " |
| 466 | "be in the interval (0, pi]." |
| 467 | ) |
| 468 | |
| 469 | if ( |
| 470 | self.recovery_angular_velocity_tolerance |
| 471 | < 0.0 |
| 472 | ): |
| 473 | raise ValueError( |
| 474 | "recovery_angular_velocity_tolerance " |
| 475 | "must not be negative." |
| 476 | ) |
| 477 | |
| 478 | if not ( |
| 479 | 0.0 |
| 480 | <= self.post_flip_shaping_gamma |
| 481 | <= 1.0 |
| 482 | ): |
| 483 | raise ValueError( |
| 484 | "post_flip_shaping_gamma must " |
| 485 | "be in [0, 1]." |
| 486 | ) |
| 487 | |
| 488 | if self.post_flip_shaping_clip <= 0.0: |
| 489 | raise ValueError( |
| 490 | "post_flip_shaping_clip must " |
| 491 | "be positive." |
| 492 | ) |
| 493 | |
| 494 | if self.pre_flip_shaping_clip <= 0.0: |
| 495 | raise ValueError( |
| 496 | "pre_flip_shaping_clip must " |
| 497 | "be positive." |
| 498 | ) |
| 499 | |
| 500 | if not ( |
| 501 | 0.0 |
| 502 | < self.flip_corridor_half_width |
| 503 | < 1.0 |
| 504 | ): |
| 505 | raise ValueError( |
| 506 | "flip_corridor_half_width must " |
| 507 | "be between 0 and 1." |
| 508 | ) |
| 509 | |
| 510 | if not ( |
| 511 | 0.0 |
| 512 | < self.landing_zone_half_width |
| 513 | < 1.0 |
| 514 | ): |
| 515 | raise ValueError( |
| 516 | "landing_zone_half_width must " |
| 517 | "be between 0 and 1." |
| 518 | ) |
| 519 | |
| 520 | if self.post_flip_max_target_vx <= 0.0: |
| 521 | raise ValueError( |
| 522 | "post_flip_max_target_vx must " |
| 523 | "be positive." |
| 524 | ) |
| 525 | |
| 526 | if self.near_ground_height <= 0.0: |
| 527 | raise ValueError( |
| 528 | "near_ground_height must be " |
| 529 | "positive." |
| 530 | ) |
| 531 | |
| 532 | # All parameters in this collection represent rewards, |
| 533 | # penalties, weights, tolerances or magnitudes. Negative values |
| 534 | # would invert their intended meaning. |
| 535 | non_negative_parameters = { |
| 536 | "pre_flip_original_reward_weight": ( |
| 537 | self.pre_flip_original_reward_weight |
| 538 | ), |
| 539 | "post_flip_original_reward_weight": ( |
| 540 | self.post_flip_original_reward_weight |
| 541 | ), |
| 542 | "pre_flip_center_weight": ( |
| 543 | self.pre_flip_center_weight |
| 544 | ), |
| 545 | "pre_flip_horizontal_speed_weight": ( |
| 546 | self.pre_flip_horizontal_speed_weight |
| 547 | ), |
| 548 | "pre_flip_shaping_weight": ( |
| 549 | self.pre_flip_shaping_weight |
| 550 | ), |
| 551 | "flip_entry_penalty_weight": ( |
| 552 | self.flip_entry_penalty_weight |
| 553 | ), |
| 554 | "rotation_progress_bonus": ( |
| 555 | self.rotation_progress_bonus |
| 556 | ), |
| 557 | "flip_completion_bonus": ( |
| 558 | self.flip_completion_bonus |
| 559 | ), |
| 560 | "recovery_bonus": ( |
| 561 | self.recovery_bonus |
| 562 | ), |
| 563 | "flip_landing_bonus": ( |
| 564 | self.flip_landing_bonus |
| 565 | ), |
| 566 | "landing_without_flip_penalty": ( |
| 567 | self.landing_without_flip_penalty |
| 568 | ), |
| 569 | "no_flip_terminal_penalty": ( |
| 570 | self.no_flip_terminal_penalty |
| 571 | ), |
| 572 | "failed_landing_penalty": ( |
| 573 | self.failed_landing_penalty |
| 574 | ), |
| 575 | "outside_zone_landing_penalty": ( |
| 576 | self.outside_zone_landing_penalty |
| 577 | ), |
| 578 | "in_zone_crash_penalty": ( |
| 579 | self.in_zone_crash_penalty |
| 580 | ), |
| 581 | "post_flip_shaping_weight": ( |
| 582 | self.post_flip_shaping_weight |
| 583 | ), |
| 584 | "post_flip_center_weight": ( |
| 585 | self.post_flip_center_weight |
| 586 | ), |
| 587 | "post_flip_horizontal_speed_weight": ( |
| 588 | self.post_flip_horizontal_speed_weight |
| 589 | ), |
| 590 | "post_flip_vertical_speed_weight": ( |
| 591 | self.post_flip_vertical_speed_weight |
| 592 | ), |
| 593 | "post_flip_angle_weight": ( |
| 594 | self.post_flip_angle_weight |
| 595 | ), |
| 596 | "post_flip_angular_speed_weight": ( |
| 597 | self.post_flip_angular_speed_weight |
| 598 | ), |
| 599 | "post_flip_leg_contact_weight": ( |
| 600 | self.post_flip_leg_contact_weight |
| 601 | ), |
| 602 | "post_flip_zone_excess_weight": ( |
| 603 | self.post_flip_zone_excess_weight |
| 604 | ), |
| 605 | "post_flip_target_vx_gain": ( |
| 606 | self.post_flip_target_vx_gain |
| 607 | ), |
| 608 | "post_flip_horizontal_deadband": ( |
| 609 | self.post_flip_horizontal_deadband |
| 610 | ), |
| 611 | "safe_touchdown_vertical_speed": ( |
| 612 | self.safe_touchdown_vertical_speed |
| 613 | ), |
| 614 | "near_ground_overspeed_weight": ( |
| 615 | self.near_ground_overspeed_weight |
| 616 | ), |
| 617 | } |
| 618 | |
| 619 | for ( |
| 620 | parameter_name, |
| 621 | parameter_value, |
| 622 | ) in non_negative_parameters.items(): |
| 623 | if parameter_value < 0.0: |
| 624 | raise ValueError( |
| 625 | f"{parameter_name} must not " |
| 626 | "be negative." |
| 627 | ) |
| 628 | |
| 629 | # Downward velocity is represented by a negative value. |
| 630 | if self.post_flip_target_vy_high > 0.0: |
| 631 | raise ValueError( |
| 632 | "post_flip_target_vy_high must " |
| 633 | "be zero or negative." |
| 634 | ) |
| 635 | |
| 636 | if ( |
| 637 | self.post_flip_target_vy_near_ground |
| 638 | > 0.0 |
| 639 | ): |
| 640 | raise ValueError( |
| 641 | "post_flip_target_vy_near_ground " |
| 642 | "must be zero or negative." |
| 643 | ) |
| 644 | |
| 645 | # For example, -0.12 is slower than -0.45. |
| 646 | if ( |
| 647 | self.post_flip_target_vy_near_ground |
| 648 | < self.post_flip_target_vy_high |
| 649 | ): |
| 650 | raise ValueError( |
| 651 | "The near-ground target vertical " |
| 652 | "speed must be slower than the " |
| 653 | "high-altitude target." |
| 654 | ) |
| 655 | |
| 656 | if ( |
| 657 | self.post_flip_horizontal_deadband |
| 658 | > self.landing_zone_half_width |
| 659 | ): |
| 660 | raise ValueError( |
| 661 | "post_flip_horizontal_deadband " |
| 662 | "must not exceed " |
| 663 | "landing_zone_half_width." |
| 664 | ) |
| 665 | |
| 666 | # Convert the configured number of rotations to radians. |
| 667 | self.required_rotation = ( |
| 668 | 2.0 |
| 669 | * math.pi |
| 670 | * self.required_rotations |
| 671 | ) |
| 672 | |
| 673 | # ============================================================== |
| 674 | # Per-episode state |
| 675 | # |
| 676 | # These values are reset at the start of every episode. They are |
| 677 | # not reward hyperparameters. |
| 678 | # ============================================================== |
| 679 | |
| 680 | self.previous_angle = 0.0 |
| 681 | self.cumulative_rotation = 0.0 |
| 682 | self.maximum_rotation_progress = 0.0 |
| 683 | self.flip_completed = False |
| 684 | self.recovery_completed = False |
| 685 | |
| 686 | self.previous_post_flip_potential: ( |
| 687 | float | None |
| 688 | ) = None |
| 689 | |
| 690 | self.previous_pre_flip_potential: ( |
| 691 | float | None |
| 692 | ) = None |
| 693 | |
| 694 | # Episode-level diagnostics. |
| 695 | self.x_at_flip_completion = np.nan |
| 696 | self.vx_at_flip_completion = np.nan |
| 697 | self.y_at_flip_completion = np.nan |
| 698 | |
| 699 | self.first_leg_contact_recorded = False |
| 700 | self.first_leg_contact_vx = np.nan |
| 701 | self.first_leg_contact_vy = np.nan |
| 702 | |
| 703 | self.total_pre_flip_shaping = 0.0 |
| 704 | self.total_post_flip_shaping = 0.0 |
| 705 | self.total_overspeed_penalty = 0.0 |
| 706 | |
| 707 | self.shaping_step_count = 0 |
| 708 | self.shaping_clip_count = 0 |
| 709 | |
| 710 | # ============================================================== |
| 711 | # Extended observation space |
| 712 | # |
| 713 | # Append: |
| 714 | # 1. signed rotation progress; |
| 715 | # 2. flip-completed flag; |
| 716 | # 3. recovery-completed flag. |
| 717 | # ============================================================== |
| 718 | |
| 719 | base_space = self.env.observation_space |
| 720 | |
| 721 | if not isinstance( |
| 722 | base_space, |
| 723 | gym.spaces.Box, |
| 724 | ): |
| 725 | raise TypeError( |
| 726 | "The wrapped environment must " |
| 727 | "have a Box observation space." |
| 728 | ) |
| 729 | |
| 730 | # Bottom bound for the 3 extra observation space parameters |
| 731 | extra_low = np.asarray( |
| 732 | [ |
| 733 | -1.0, |
| 734 | 0.0, |
| 735 | 0.0, |
| 736 | ], |
| 737 | dtype=np.float32, |
| 738 | ) |
| 739 | |
| 740 | # Top bound for the 3 extra observation space parameters |
| 741 | extra_high = np.asarray( |
| 742 | [ |
| 743 | 1.0, |
| 744 | 1.0, |
| 745 | 1.0, |
| 746 | ], |
| 747 | dtype=np.float32, |
| 748 | ) |
| 749 | |
| 750 | self.observation_space = ( |
| 751 | gym.spaces.Box( |
| 752 | low=np.concatenate( |
| 753 | [ |
| 754 | base_space.low.astype( |
| 755 | np.float32 |
| 756 | ), |
| 757 | extra_low, |
| 758 | ] |
| 759 | ), |
| 760 | high=np.concatenate( |
| 761 | [ |
| 762 | base_space.high.astype( |
| 763 | np.float32 |
| 764 | ), |
| 765 | extra_high, |
| 766 | ] |
| 767 | ), |
| 768 | dtype=np.float32, |
| 769 | ) |
| 770 | ) |
| 771 | |
| 772 | @staticmethod |
| 773 | def _wrap_angle(angle: float) -> float: |
| 774 | """Wrap an angle to [-pi, pi].""" |
| 775 | |
| 776 | return float( |
| 777 | np.arctan2( |
| 778 | np.sin(angle), |
| 779 | np.cos(angle), |
| 780 | ) |
| 781 | ) |
| 782 | |
| 783 | def _signed_rotation_progress(self) -> float: |
| 784 | return float( |
| 785 | np.clip( |
| 786 | self.cumulative_rotation |
| 787 | / self.required_rotation, |
| 788 | -1.0, |
| 789 | 1.0, |
| 790 | ) |
| 791 | ) |
| 792 | |
| 793 | def _augment_observation( |
| 794 | self, |
| 795 | observation: np.ndarray, |
| 796 | ) -> np.ndarray: |
| 797 | """ |
| 798 | Adds the custom recorded values to the original observation. |
| 799 | """ |
| 800 | return np.concatenate( |
| 801 | [ |
| 802 | np.asarray( |
| 803 | observation, |
| 804 | dtype=np.float32, |
| 805 | ), |
| 806 | np.asarray( |
| 807 | [ |
| 808 | self._signed_rotation_progress(), |
| 809 | float(self.flip_completed), |
| 810 | float(self.recovery_completed), |
| 811 | ], |
| 812 | dtype=np.float32, |
| 813 | ), |
| 814 | ] |
| 815 | ) |
| 816 | |
| 817 | def _pre_flip_potential( |
| 818 | self, |
| 819 | observation: np.ndarray, |
| 820 | ) -> float: |
| 821 | """ |
| 822 | Higher values represent better pre-flip horizontal states. |
| 823 | |
| 824 | Only horizontal position and horizontal speed are used. |
| 825 | Angle is deliberately excluded so this shaping does not |
| 826 | oppose the required rotation. |
| 827 | """ |
| 828 | |
| 829 | x_position = float( |
| 830 | observation[0] |
| 831 | ) |
| 832 | |
| 833 | horizontal_speed = float( |
| 834 | observation[2] |
| 835 | ) |
| 836 | |
| 837 | return float( |
| 838 | -self.pre_flip_center_weight |
| 839 | * abs(x_position) |
| 840 | |
| 841 | -self.pre_flip_horizontal_speed_weight |
| 842 | * abs(horizontal_speed) |
| 843 | ) |
| 844 | |
| 845 | def _post_flip_potential( |
| 846 | self, |
| 847 | observation: np.ndarray, |
| 848 | ) -> float: |
| 849 | """ |
| 850 | Higher values represent safer post-flip states. |
| 851 | |
| 852 | LunarLander-v3 observation positions: |
| 853 | 0 x position, 2 x velocity, 4 angle, 5 angular velocity, |
| 854 | 6 left-leg contact and 7 right-leg contact. |
| 855 | """ |
| 856 | |
| 857 | x_position = float( |
| 858 | observation[0] |
| 859 | ) |
| 860 | |
| 861 | y_position = float( |
| 862 | observation[1] |
| 863 | ) |
| 864 | |
| 865 | horizontal_speed = float( |
| 866 | observation[2] |
| 867 | ) |
| 868 | |
| 869 | vertical_speed = float( |
| 870 | observation[3] |
| 871 | ) |
| 872 | |
| 873 | height_ratio = float( |
| 874 | np.clip( |
| 875 | y_position |
| 876 | / self.near_ground_height, |
| 877 | 0.0, |
| 878 | 1.0, |
| 879 | ) |
| 880 | ) |
| 881 | |
| 882 | target_vertical_speed = ( |
| 883 | self.post_flip_target_vy_near_ground |
| 884 | + height_ratio |
| 885 | * ( |
| 886 | self.post_flip_target_vy_high |
| 887 | - self.post_flip_target_vy_near_ground |
| 888 | ) |
| 889 | ) |
| 890 | |
| 891 | vertical_speed_error = abs( |
| 892 | vertical_speed |
| 893 | - target_vertical_speed |
| 894 | ) |
| 895 | |
| 896 | wrapped_angle = abs( |
| 897 | self._wrap_angle( |
| 898 | float(observation[4]) |
| 899 | ) |
| 900 | ) |
| 901 | |
| 902 | angular_speed = abs( |
| 903 | float(observation[5]) |
| 904 | ) |
| 905 | |
| 906 | leg_contacts = ( |
| 907 | float(observation[6] > 0.5) |
| 908 | + float(observation[7] > 0.5) |
| 909 | ) |
| 910 | |
| 911 | # Distance beyond the two landing-zone flags. |
| 912 | zone_excess = max( |
| 913 | abs(x_position) |
| 914 | - self.landing_zone_half_width, |
| 915 | 0.0, |
| 916 | ) |
| 917 | |
| 918 | # When left of the pad, target positive horizontal |
| 919 | # velocity. When right of the pad, target negative |
| 920 | # horizontal velocity. Close to the centre, target |
| 921 | # velocity approaches zero. |
| 922 | horizontal_guidance_factor = float( |
| 923 | np.clip( |
| 924 | y_position / 0.50, |
| 925 | 0.0, |
| 926 | 1.0, |
| 927 | ) |
| 928 | ) |
| 929 | |
| 930 | horizontal_error = max( |
| 931 | abs(x_position) |
| 932 | - self.post_flip_horizontal_deadband, |
| 933 | 0.0, |
| 934 | ) |
| 935 | |
| 936 | target_horizontal_speed = float( |
| 937 | np.clip( |
| 938 | -self.post_flip_target_vx_gain |
| 939 | * x_position |
| 940 | * horizontal_guidance_factor, |
| 941 | -self.post_flip_max_target_vx, |
| 942 | self.post_flip_max_target_vx, |
| 943 | ) |
| 944 | ) |
| 945 | |
| 946 | horizontal_speed_error = abs( |
| 947 | horizontal_speed |
| 948 | - target_horizontal_speed |
| 949 | ) |
| 950 | |
| 951 | # Increase the importance of entering the landing |
| 952 | # zone as the lander approaches the ground. |
| 953 | near_ground_factor = float( |
| 954 | np.clip( |
| 955 | 1.0 |
| 956 | - max(y_position, 0.0) |
| 957 | / 0.75, |
| 958 | 0.0, |
| 959 | 1.0, |
| 960 | ) |
| 961 | ) |
| 962 | |
| 963 | effective_zone_weight = ( |
| 964 | self.post_flip_zone_excess_weight |
| 965 | * (1.0 + near_ground_factor) |
| 966 | ) |
| 967 | |
| 968 | return float( |
| 969 | -self.post_flip_center_weight |
| 970 | * horizontal_error |
| 971 | |
| 972 | -effective_zone_weight |
| 973 | * zone_excess |
| 974 | |
| 975 | -self.post_flip_horizontal_speed_weight |
| 976 | * horizontal_speed_error |
| 977 | |
| 978 | -self.post_flip_vertical_speed_weight |
| 979 | * vertical_speed_error |
| 980 | |
| 981 | -self.post_flip_angle_weight |
| 982 | * wrapped_angle |
| 983 | |
| 984 | -self.post_flip_angular_speed_weight |
| 985 | * angular_speed |
| 986 | |
| 987 | +self.post_flip_leg_contact_weight |
| 988 | * leg_contacts |
| 989 | ) |
| 990 | |
| 991 | def reset( |
| 992 | self, |
| 993 | *, |
| 994 | seed: int | None = None, |
| 995 | options: dict | None = None, |
| 996 | ): |
| 997 | observation, info = self.env.reset( |
| 998 | seed=seed, |
| 999 | options=options, |
| 1000 | ) |
| 1001 | |
| 1002 | self.previous_angle = float(observation[4]) |
| 1003 | self.cumulative_rotation = 0.0 |
| 1004 | self.maximum_rotation_progress = 0.0 |
| 1005 | self.flip_completed = False |
| 1006 | self.recovery_completed = False |
| 1007 | self.previous_post_flip_potential = None |
| 1008 | self.previous_pre_flip_potential = ( |
| 1009 | self._pre_flip_potential( |
| 1010 | observation |
| 1011 | ) |
| 1012 | ) |
| 1013 | |
| 1014 | self.x_at_flip_completion = np.nan |
| 1015 | self.vx_at_flip_completion = np.nan |
| 1016 | self.y_at_flip_completion = np.nan |
| 1017 | |
| 1018 | self.first_leg_contact_recorded = False |
| 1019 | self.first_leg_contact_vx = np.nan |
| 1020 | self.first_leg_contact_vy = np.nan |
| 1021 | |
| 1022 | self.total_pre_flip_shaping = 0.0 |
| 1023 | self.total_post_flip_shaping = 0.0 |
| 1024 | self.total_overspeed_penalty = 0.0 |
| 1025 | |
| 1026 | self.shaping_step_count = 0 |
| 1027 | self.shaping_clip_count = 0 |
| 1028 | |
| 1029 | info = dict(info) |
| 1030 | info.update( |
| 1031 | { |
| 1032 | "flip_completed": False, |
| 1033 | "recovery_completed": False, |
| 1034 | "rotation_progress": 0.0, |
| 1035 | "rotations_completed": 0.0, |
| 1036 | "rotation_progress_reward": 0.0, |
| 1037 | "flip_completion_reward": 0.0, |
| 1038 | "recovery_reward": 0.0, |
| 1039 | "post_flip_shaping_reward": 0.0, |
| 1040 | |
| 1041 | "pre_flip_shaping_reward": 0.0, |
| 1042 | "flip_entry_penalty": 0.0, |
| 1043 | |
| 1044 | "x_at_flip_completion": np.nan, |
| 1045 | "vx_at_flip_completion": np.nan, |
| 1046 | "y_at_flip_completion": np.nan, |
| 1047 | |
| 1048 | "first_leg_contact_vx": np.nan, |
| 1049 | "first_leg_contact_vy": np.nan, |
| 1050 | |
| 1051 | "total_pre_flip_shaping": 0.0, |
| 1052 | "total_post_flip_shaping": 0.0, |
| 1053 | "total_overspeed_penalty": 0.0, |
| 1054 | |
| 1055 | "shaping_step_count": 0, |
| 1056 | "shaping_clip_count": 0, |
| 1057 | "shaping_clip_rate": 0.0, |
| 1058 | |
| 1059 | "flip_landing_bonus": 0.0, |
| 1060 | "terminal_adjustment": 0.0, |
| 1061 | } |
| 1062 | ) |
| 1063 | |
| 1064 | return self._augment_observation(observation), info |
| 1065 | |
| 1066 | def step(self, action): |
| 1067 | # 1. Advance the original LunarLander environment. |
| 1068 | ( |
| 1069 | observation, |
| 1070 | original_reward, |
| 1071 | terminated, |
| 1072 | truncated, |
| 1073 | info, |
| 1074 | ) = self.env.step(action) |
| 1075 | |
| 1076 | # Remember whether the flip had already been completed |
| 1077 | # before processing the current transition. |
| 1078 | flip_was_completed = self.flip_completed |
| 1079 | |
| 1080 | # 2. Accumulate directed angular movement and reward new flip progress. |
| 1081 | current_angle = float(observation[4]) |
| 1082 | angular_velocity = float(observation[5]) |
| 1083 | |
| 1084 | angle_change = self._wrap_angle( |
| 1085 | current_angle - self.previous_angle |
| 1086 | ) |
| 1087 | self.previous_angle = current_angle |
| 1088 | |
| 1089 | directed_change = ( |
| 1090 | self.rotation_direction * angle_change |
| 1091 | ) |
| 1092 | self.cumulative_rotation += directed_change |
| 1093 | |
| 1094 | rotation_progress = float( |
| 1095 | np.clip( |
| 1096 | self.cumulative_rotation |
| 1097 | / self.required_rotation, |
| 1098 | 0.0, |
| 1099 | 1.0, |
| 1100 | ) |
| 1101 | ) |
| 1102 | |
| 1103 | new_progress = max( |
| 1104 | 0.0, |
| 1105 | rotation_progress |
| 1106 | - self.maximum_rotation_progress, |
| 1107 | ) |
| 1108 | progress_reward = ( |
| 1109 | self.rotation_progress_bonus |
| 1110 | * new_progress |
| 1111 | ) |
| 1112 | self.maximum_rotation_progress = max( |
| 1113 | self.maximum_rotation_progress, |
| 1114 | rotation_progress, |
| 1115 | ) |
| 1116 | |
| 1117 | # Apply horizontal guidance before the flip is completed. |
| 1118 | # This rewards becoming more centred and reducing lateral |
| 1119 | # speed without applying any pressure to remain upright. |
| 1120 | pre_flip_shaping_reward = 0.0 |
| 1121 | |
| 1122 | if ( |
| 1123 | not flip_was_completed |
| 1124 | and self.pre_flip_shaping_weight > 0.0 |
| 1125 | ): |
| 1126 | current_pre_flip_potential = ( |
| 1127 | self._pre_flip_potential( |
| 1128 | observation |
| 1129 | ) |
| 1130 | ) |
| 1131 | |
| 1132 | if ( |
| 1133 | self.previous_pre_flip_potential |
| 1134 | is not None |
| 1135 | ): |
| 1136 | raw_pre_flip_shaping = ( |
| 1137 | self.pre_flip_shaping_weight |
| 1138 | * ( |
| 1139 | self.post_flip_shaping_gamma |
| 1140 | * current_pre_flip_potential |
| 1141 | - self.previous_pre_flip_potential |
| 1142 | ) |
| 1143 | ) |
| 1144 | |
| 1145 | pre_flip_shaping_reward = float( |
| 1146 | np.clip( |
| 1147 | raw_pre_flip_shaping, |
| 1148 | -self.pre_flip_shaping_clip, |
| 1149 | self.pre_flip_shaping_clip, |
| 1150 | ) |
| 1151 | ) |
| 1152 | |
| 1153 | self.shaping_step_count += 1 |
| 1154 | |
| 1155 | if ( |
| 1156 | abs(raw_pre_flip_shaping) |
| 1157 | >= self.pre_flip_shaping_clip |
| 1158 | ): |
| 1159 | self.shaping_clip_count += 1 |
| 1160 | |
| 1161 | self.previous_pre_flip_potential = ( |
| 1162 | current_pre_flip_potential |
| 1163 | ) |
| 1164 | |
| 1165 | completion_reward = 0.0 |
| 1166 | flip_entry_penalty = 0.0 |
| 1167 | |
| 1168 | # 3. Detect the first completed full rotation. |
| 1169 | if ( |
| 1170 | rotation_progress >= 1.0 |
| 1171 | and not self.flip_completed |
| 1172 | ): |
| 1173 | self.flip_completed = True |
| 1174 | completion_reward = ( |
| 1175 | self.flip_completion_bonus |
| 1176 | ) |
| 1177 | |
| 1178 | self.x_at_flip_completion = float( |
| 1179 | observation[0] |
| 1180 | ) |
| 1181 | |
| 1182 | self.vx_at_flip_completion = float( |
| 1183 | observation[2] |
| 1184 | ) |
| 1185 | |
| 1186 | self.y_at_flip_completion = float( |
| 1187 | observation[1] |
| 1188 | ) |
| 1189 | |
| 1190 | off_center_at_flip = max( |
| 1191 | abs(self.x_at_flip_completion) |
| 1192 | - self.flip_corridor_half_width, |
| 1193 | 0.0, |
| 1194 | ) |
| 1195 | |
| 1196 | flip_entry_penalty = ( |
| 1197 | self.flip_entry_penalty_weight |
| 1198 | * off_center_at_flip |
| 1199 | ) |
| 1200 | |
| 1201 | # 4. Detect upright post-flip recovery. |
| 1202 | wrapped_angle = abs( |
| 1203 | self._wrap_angle(current_angle) |
| 1204 | ) |
| 1205 | upright = ( |
| 1206 | wrapped_angle |
| 1207 | <= self.upright_tolerance_radians |
| 1208 | ) |
| 1209 | |
| 1210 | recovery_reward = 0.0 |
| 1211 | |
| 1212 | if ( |
| 1213 | self.flip_completed |
| 1214 | and not self.recovery_completed |
| 1215 | and upright |
| 1216 | and abs(angular_velocity) |
| 1217 | <= self.recovery_angular_velocity_tolerance |
| 1218 | ): |
| 1219 | self.recovery_completed = True |
| 1220 | recovery_reward = self.recovery_bonus |
| 1221 | |
| 1222 | |
| 1223 | post_flip_shaping_reward = 0.0 |
| 1224 | post_flip_potential = None |
| 1225 | near_ground_overspeed_penalty = 0.0 |
| 1226 | |
| 1227 | # 5. Apply an explicit near-ground descent-overspeed penalty. |
| 1228 | if self.flip_completed: |
| 1229 | y_position = float( |
| 1230 | observation[1] |
| 1231 | ) |
| 1232 | |
| 1233 | vertical_speed = float( |
| 1234 | observation[3] |
| 1235 | ) |
| 1236 | |
| 1237 | near_ground_factor = float( |
| 1238 | np.clip( |
| 1239 | 1.0 |
| 1240 | - max(y_position, 0.0) |
| 1241 | / self.near_ground_height, |
| 1242 | 0.0, |
| 1243 | 1.0, |
| 1244 | ) |
| 1245 | ) |
| 1246 | |
| 1247 | downward_speed = max( |
| 1248 | -vertical_speed, |
| 1249 | 0.0, |
| 1250 | ) |
| 1251 | |
| 1252 | descent_overspeed = max( |
| 1253 | downward_speed |
| 1254 | - self.safe_touchdown_vertical_speed, |
| 1255 | 0.0, |
| 1256 | ) |
| 1257 | |
| 1258 | near_ground_overspeed_penalty = ( |
| 1259 | self.near_ground_overspeed_weight |
| 1260 | * near_ground_factor |
| 1261 | * descent_overspeed**2 |
| 1262 | ) |
| 1263 | |
| 1264 | # 6. Apply continuous post-flip landing-quality shaping. |
| 1265 | if self.flip_completed: |
| 1266 | post_flip_potential = ( |
| 1267 | self._post_flip_potential(observation) |
| 1268 | ) |
| 1269 | |
| 1270 | if self.previous_post_flip_potential is None: |
| 1271 | self.previous_post_flip_potential = ( |
| 1272 | post_flip_potential |
| 1273 | ) |
| 1274 | else: |
| 1275 | potential_difference = ( |
| 1276 | self.post_flip_shaping_gamma |
| 1277 | * post_flip_potential |
| 1278 | - self.previous_post_flip_potential |
| 1279 | ) |
| 1280 | |
| 1281 | raw_post_flip_shaping = ( |
| 1282 | self.post_flip_shaping_weight |
| 1283 | * potential_difference |
| 1284 | ) |
| 1285 | |
| 1286 | post_flip_shaping_reward = float( |
| 1287 | np.clip( |
| 1288 | raw_post_flip_shaping, |
| 1289 | -self.post_flip_shaping_clip, |
| 1290 | self.post_flip_shaping_clip, |
| 1291 | ) |
| 1292 | ) |
| 1293 | |
| 1294 | self.shaping_step_count += 1 |
| 1295 | |
| 1296 | if ( |
| 1297 | abs(raw_post_flip_shaping) |
| 1298 | >= self.post_flip_shaping_clip |
| 1299 | ): |
| 1300 | self.shaping_clip_count += 1 |
| 1301 | |
| 1302 | self.previous_post_flip_potential = ( |
| 1303 | post_flip_potential |
| 1304 | ) |
| 1305 | |
| 1306 | left_leg_contact = bool(observation[6] > 0.5) |
| 1307 | right_leg_contact = bool(observation[7] > 0.5) |
| 1308 | |
| 1309 | if ( |
| 1310 | self.flip_completed |
| 1311 | and not self.first_leg_contact_recorded |
| 1312 | and ( |
| 1313 | left_leg_contact |
| 1314 | or right_leg_contact |
| 1315 | ) |
| 1316 | ): |
| 1317 | self.first_leg_contact_recorded = True |
| 1318 | |
| 1319 | self.first_leg_contact_vx = float( |
| 1320 | observation[2] |
| 1321 | ) |
| 1322 | |
| 1323 | self.first_leg_contact_vy = float( |
| 1324 | observation[3] |
| 1325 | ) |
| 1326 | |
| 1327 | x_position = float( |
| 1328 | observation[0] |
| 1329 | ) |
| 1330 | |
| 1331 | inside_landing_zone = bool( |
| 1332 | abs(x_position) |
| 1333 | <= self.landing_zone_half_width |
| 1334 | ) |
| 1335 | |
| 1336 | stable_landing = bool( |
| 1337 | terminated |
| 1338 | and float(original_reward) > 0.0 |
| 1339 | and left_leg_contact |
| 1340 | and right_leg_contact |
| 1341 | and upright |
| 1342 | ) |
| 1343 | |
| 1344 | # "Safe landing" now means stable and |
| 1345 | # between the landing-zone boundaries. |
| 1346 | landed_safely = bool( |
| 1347 | stable_landing |
| 1348 | and inside_landing_zone |
| 1349 | ) |
| 1350 | |
| 1351 | episode_finished = bool( |
| 1352 | terminated or truncated |
| 1353 | ) |
| 1354 | |
| 1355 | # 7. Classify the terminal outcome and apply its one-off adjustment. |
| 1356 | terminal_adjustment = 0.0 |
| 1357 | outcome = "in_progress" |
| 1358 | |
| 1359 | # Complete success: |
| 1360 | # flipped, settled safely and landed inside the zone. |
| 1361 | if ( |
| 1362 | self.flip_completed |
| 1363 | and landed_safely |
| 1364 | ): |
| 1365 | terminal_adjustment = ( |
| 1366 | self.flip_landing_bonus |
| 1367 | ) |
| 1368 | |
| 1369 | outcome = ( |
| 1370 | "flip_and_safe_landing" |
| 1371 | ) |
| 1372 | |
| 1373 | elif episode_finished: |
| 1374 | # The episode ended before completing the flip. |
| 1375 | if not self.flip_completed: |
| 1376 | if stable_landing: |
| 1377 | terminal_adjustment = -( |
| 1378 | self.landing_without_flip_penalty |
| 1379 | ) |
| 1380 | |
| 1381 | outcome = ( |
| 1382 | "stable_landing_without_flip" |
| 1383 | ) |
| 1384 | |
| 1385 | else: |
| 1386 | terminal_adjustment = -( |
| 1387 | self.no_flip_terminal_penalty |
| 1388 | ) |
| 1389 | |
| 1390 | outcome = ( |
| 1391 | "episode_ended_without_flip" |
| 1392 | ) |
| 1393 | |
| 1394 | # Step E: |
| 1395 | # The flip was completed and the lander reached |
| 1396 | # the zone, but it crashed instead of settling. |
| 1397 | elif ( |
| 1398 | terminated |
| 1399 | and inside_landing_zone |
| 1400 | and not stable_landing |
| 1401 | ): |
| 1402 | terminal_adjustment = -( |
| 1403 | self.in_zone_crash_penalty |
| 1404 | ) |
| 1405 | |
| 1406 | outcome = ( |
| 1407 | "flip_and_crash_in_zone" |
| 1408 | ) |
| 1409 | |
| 1410 | # A stable landing occurred, but outside the flags. |
| 1411 | elif ( |
| 1412 | stable_landing |
| 1413 | and not inside_landing_zone |
| 1414 | ): |
| 1415 | terminal_adjustment = -( |
| 1416 | self.outside_zone_landing_penalty |
| 1417 | ) |
| 1418 | |
| 1419 | outcome = ( |
| 1420 | "flip_and_stable_landing_" |
| 1421 | "outside_zone" |
| 1422 | ) |
| 1423 | |
| 1424 | # Other post-flip failures, such as crashing |
| 1425 | # outside the zone or timing out. |
| 1426 | else: |
| 1427 | terminal_adjustment = -( |
| 1428 | self.failed_landing_penalty |
| 1429 | ) |
| 1430 | |
| 1431 | outcome = ( |
| 1432 | "flip_but_failed_landing" |
| 1433 | ) |
| 1434 | |
| 1435 | self.total_pre_flip_shaping += ( |
| 1436 | pre_flip_shaping_reward |
| 1437 | ) |
| 1438 | |
| 1439 | self.total_post_flip_shaping += ( |
| 1440 | post_flip_shaping_reward |
| 1441 | ) |
| 1442 | |
| 1443 | self.total_overspeed_penalty += ( |
| 1444 | near_ground_overspeed_penalty |
| 1445 | ) |
| 1446 | |
| 1447 | shaping_clip_rate = ( |
| 1448 | self.shaping_clip_count |
| 1449 | / self.shaping_step_count |
| 1450 | if self.shaping_step_count > 0 |
| 1451 | else 0.0 |
| 1452 | ) |
| 1453 | |
| 1454 | original_weight = ( |
| 1455 | self.post_flip_original_reward_weight |
| 1456 | if self.flip_completed |
| 1457 | else self.pre_flip_original_reward_weight |
| 1458 | ) |
| 1459 | |
| 1460 | # 8. Combine the original reward, dense shaping and terminal outcome. |
| 1461 | modified_reward = ( |
| 1462 | original_weight * float(original_reward) |
| 1463 | + progress_reward |
| 1464 | + pre_flip_shaping_reward |
| 1465 | + completion_reward |
| 1466 | + recovery_reward |
| 1467 | + post_flip_shaping_reward |
| 1468 | - flip_entry_penalty |
| 1469 | - near_ground_overspeed_penalty |
| 1470 | + terminal_adjustment |
| 1471 | ) |
| 1472 | |
| 1473 | info = dict(info) |
| 1474 | info.update( |
| 1475 | { |
| 1476 | "original_reward": float(original_reward), |
| 1477 | "rotation_progress": rotation_progress, |
| 1478 | "rotation_progress_reward": progress_reward, |
| 1479 | "flip_completion_reward": completion_reward, |
| 1480 | "recovery_reward": recovery_reward, |
| 1481 | "post_flip_shaping_reward": ( |
| 1482 | post_flip_shaping_reward |
| 1483 | ), |
| 1484 | "post_flip_potential": post_flip_potential, |
| 1485 | "cumulative_rotation": ( |
| 1486 | self.cumulative_rotation |
| 1487 | ), |
| 1488 | "rotations_completed": ( |
| 1489 | self.cumulative_rotation |
| 1490 | / (2.0 * math.pi) |
| 1491 | ), |
| 1492 | "flip_completed": self.flip_completed, |
| 1493 | "recovery_completed": ( |
| 1494 | self.recovery_completed |
| 1495 | ), |
| 1496 | "landed_safely": landed_safely, |
| 1497 | "original_reward_weight": original_weight, |
| 1498 | "terminal_adjustment": ( |
| 1499 | terminal_adjustment |
| 1500 | ), |
| 1501 | "flip_landing_bonus": ( |
| 1502 | self.flip_landing_bonus |
| 1503 | if outcome |
| 1504 | == "flip_and_safe_landing" |
| 1505 | else 0.0 |
| 1506 | ), |
| 1507 | "custom_outcome": outcome, |
| 1508 | "stable_landing": stable_landing, |
| 1509 | "inside_landing_zone": ( |
| 1510 | inside_landing_zone |
| 1511 | ), |
| 1512 | "horizontal_position": ( |
| 1513 | x_position |
| 1514 | ), |
| 1515 | "near_ground_overspeed_penalty": ( |
| 1516 | near_ground_overspeed_penalty |
| 1517 | ), |
| 1518 | "vertical_speed": float( |
| 1519 | observation[3] |
| 1520 | ), |
| 1521 | |
| 1522 | "pre_flip_shaping_reward": ( |
| 1523 | pre_flip_shaping_reward |
| 1524 | ), |
| 1525 | "flip_entry_penalty": ( |
| 1526 | flip_entry_penalty |
| 1527 | ), |
| 1528 | |
| 1529 | "x_at_flip_completion": ( |
| 1530 | self.x_at_flip_completion |
| 1531 | ), |
| 1532 | "vx_at_flip_completion": ( |
| 1533 | self.vx_at_flip_completion |
| 1534 | ), |
| 1535 | "y_at_flip_completion": ( |
| 1536 | self.y_at_flip_completion |
| 1537 | ), |
| 1538 | |
| 1539 | "first_leg_contact_vx": ( |
| 1540 | self.first_leg_contact_vx |
| 1541 | ), |
| 1542 | "first_leg_contact_vy": ( |
| 1543 | self.first_leg_contact_vy |
| 1544 | ), |
| 1545 | |
| 1546 | "total_pre_flip_shaping": ( |
| 1547 | self.total_pre_flip_shaping |
| 1548 | ), |
| 1549 | "total_post_flip_shaping": ( |
| 1550 | self.total_post_flip_shaping |
| 1551 | ), |
| 1552 | "total_overspeed_penalty": ( |
| 1553 | self.total_overspeed_penalty |
| 1554 | ), |
| 1555 | |
| 1556 | "shaping_step_count": ( |
| 1557 | self.shaping_step_count |
| 1558 | ), |
| 1559 | "shaping_clip_count": ( |
| 1560 | self.shaping_clip_count |
| 1561 | ), |
| 1562 | "shaping_clip_rate": ( |
| 1563 | shaping_clip_rate |
| 1564 | ), |
| 1565 | } |
| 1566 | ) |
| 1567 | |
| 1568 | return ( |
| 1569 | self._augment_observation(observation), |
| 1570 | modified_reward, |
| 1571 | terminated, |
| 1572 | truncated, |
| 1573 | info, |
| 1574 | ) |
| 1575 | |
| 1576 | |
| 1577 | def make_flip_lander( |
| 1578 | *, |
| 1579 | render_mode: str | None = None, |
| 1580 | reward_config: ( |
| 1581 | Mapping[str, float | int] |
| 1582 | | None |
| 1583 | ) = None, |
| 1584 | ) -> gym.Env: |
| 1585 | """ |
| 1586 | Build the exact environment used for training and evaluation. |
| 1587 | |
| 1588 | Pass the same reward_config to the uploader so the Hub model card |
| 1589 | records the environment accurately. |
| 1590 | """ |
| 1591 | |
| 1592 | base_env = gym.make( |
| 1593 | "LunarLander-v3", |
| 1594 | render_mode=render_mode, |
| 1595 | ) |
| 1596 | |
| 1597 | return FlipLandingRewardWrapper( |
| 1598 | base_env, |
| 1599 | reward_config=reward_config, |
| 1600 | ) |
| 1601 | |