File size: 72,352 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 |
{
"paper_id": "2022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:13:20.477723Z"
},
"title": "On the Limits of Evaluating Embodied Agent Model Generalization Using Validation Sets",
"authors": [
{
"first": "Hyounghun",
"middle": [],
"last": "Kim",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "UNC Chapel Hill",
"location": {}
},
"email": "hyounghk@cs.unc.edu"
},
{
"first": "Aishwarya",
"middle": [],
"last": "Padmakumar",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Amazon Alexa AI",
"location": {}
},
"email": "padmakua@amazon.com"
},
{
"first": "Di",
"middle": [],
"last": "Jin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Amazon Alexa AI",
"location": {}
},
"email": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Bansal",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "UNC Chapel Hill",
"location": {}
},
"email": "mbansal@cs.unc.edu"
},
{
"first": "Dilek",
"middle": [],
"last": "Hakkani-Tur",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Amazon Alexa AI",
"location": {}
},
"email": "hakkanit@amazon.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Natural language guided embodied task completion is a challenging problem since it requires understanding natural language instructions, aligning them with egocentric visual observations, and choosing appropriate actions to execute in the environment to produce desired changes. We experiment with augmenting a transformer model for this task with modules that effectively utilize a wider field of view and learn to choose whether the next step requires a navigation or manipulation action. We observed that the proposed modules resulted in improved, and in fact state-of-the-art performance on an unseen validation set of a popular benchmark dataset, ALFRED. However, our best model selected using the unseen validation set underperforms on the unseen test split of ALFRED, indicating that performance on the unseen validation set may not in itself be a sufficient indicator of whether model improvements generalize to unseen test sets. We highlight this result as we believe it may be a wider phenomenon in machine learning tasks but primarily noticeable only in benchmarks that limit evaluations on test splits, and highlights the need to modify benchmark design to better account for variance in model performance.",
"pdf_parse": {
"paper_id": "2022",
"_pdf_hash": "",
"abstract": [
{
"text": "Natural language guided embodied task completion is a challenging problem since it requires understanding natural language instructions, aligning them with egocentric visual observations, and choosing appropriate actions to execute in the environment to produce desired changes. We experiment with augmenting a transformer model for this task with modules that effectively utilize a wider field of view and learn to choose whether the next step requires a navigation or manipulation action. We observed that the proposed modules resulted in improved, and in fact state-of-the-art performance on an unseen validation set of a popular benchmark dataset, ALFRED. However, our best model selected using the unseen validation set underperforms on the unseen test split of ALFRED, indicating that performance on the unseen validation set may not in itself be a sufficient indicator of whether model improvements generalize to unseen test sets. We highlight this result as we believe it may be a wider phenomenon in machine learning tasks but primarily noticeable only in benchmarks that limit evaluations on test splits, and highlights the need to modify benchmark design to better account for variance in model performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Language guided embodied task completion is an important skill for embodied agents requiring them to follow natural language instructions to navigate in their environment and manipulate objects to complete tasks. Natural language is an easy medium for users to interact with embodied agents and effective use of natural language instructions can enable agents to navigate more easily in previously unexplored environments, and complete tasks involving novel combinations of object manipulations. Vision and language navigation benchmarks (Anderson et al., 2018; Thomason et al., 2019; Ku et al., 2020) provide an agent with natural language route instructions and evaluate their ability to follow these to navigate to a target location. It requires agents to have a deep understanding of natural language instructions, ground these in egocentric image observations and predict a sequence of actions in the environment. Other benchmarks study the manipulation and arrangement of objects (Bisk et al., 2016; Wang et al., 2016; Li et al., 2016; Bisk et al., 2018) -another crucial skill to complete many tasks that users may desire embodied agents to be able to complete. These tasks additionally require agents to reason about the states of objects and relations between them. Language guided embodied task completion benchmarks (Shridhar et al., 2020; Kim et al., 2020; Padmakumar et al., 2022) combine these skills -requiring agents to perform both navigation and object manipulation/arrangement following natural language instructions.",
"cite_spans": [
{
"start": 538,
"end": 561,
"text": "(Anderson et al., 2018;",
"ref_id": "BIBREF0"
},
{
"start": 562,
"end": 584,
"text": "Thomason et al., 2019;",
"ref_id": "BIBREF18"
},
{
"start": 585,
"end": 601,
"text": "Ku et al., 2020)",
"ref_id": "BIBREF8"
},
{
"start": 986,
"end": 1005,
"text": "(Bisk et al., 2016;",
"ref_id": "BIBREF1"
},
{
"start": 1006,
"end": 1024,
"text": "Wang et al., 2016;",
"ref_id": "BIBREF19"
},
{
"start": 1025,
"end": 1041,
"text": "Li et al., 2016;",
"ref_id": "BIBREF9"
},
{
"start": 1042,
"end": 1060,
"text": "Bisk et al., 2018)",
"ref_id": "BIBREF2"
},
{
"start": 1327,
"end": 1350,
"text": "(Shridhar et al., 2020;",
"ref_id": "BIBREF16"
},
{
"start": 1351,
"end": 1368,
"text": "Kim et al., 2020;",
"ref_id": "BIBREF6"
},
{
"start": 1369,
"end": 1393,
"text": "Padmakumar et al., 2022)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we explore a challenging navigation and manipulation benchmark, AL-FRED (Shridhar et al., 2020) , where an agent has to learn to follow complex hierarchical natural language instructions to complete tasks by navigating in a virtual environment and manipulating objects to produce desired state changes. The ALFRED benchmark provides a training dataset of action trajectories taken by an embodied agent in a variety of simulated indoor rooms paired with hierarchical natural language instructions describing the task to be accomplished and the steps to be taken to do so. For validation and testing of models, there are two splits each -seen and unseen splits. The seen validation and testing splits consist of instructions set in the same rooms as those in the training set, while the unseen splits consist of instructions set in rooms the agent has never seen before, with rooms in the unseen test set being different from those in the train and unseen validation set. Performance on the unseen validation and test sets are considered to be the best indicators of whether a model can really solve the task as the agent must operate in a completely novel floorplan, and cannot rely for example on memorized locations of large objects such as a fridge or a sink. Additionally, the ground truth action sequences are not publicly available for the seen and unseen test sets, and participants must submit prediction acted sequences on the test sets to an evaluation server where they are privately evaluated to obtain test performance. The evaluation server limits the number of submissions that can be made from an account to one per week to discourage directly tuning hyperparameters of a model on the test set. It is expected that following standard procedure in training machine learning models, one may use the validation sets to evaluate models trained with different hyperparameters, or ablating different components on the validation sets and only evaluate the best model on the test sets. Since ideally we would want a model to perform well on the unseen test set, it is reasonable to use success rate on the unseen validation set as a metric to choose which model is to be submitted for evaluation on the unseen test set.",
"cite_spans": [
{
"start": 87,
"end": 110,
"text": "(Shridhar et al., 2020)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One technique previously demonstrated to improve performance on ALFRED is the use of a multi-view setup (Nguyen et al., 2021; Kim et al., 2021) where an agent turns or moves its head in place at every time step to obtain additional views before deciding what action to take. In contrast to current models that simply concatenate features from each view, we use view-action matching -explicitly aligning embeddings of actions with embeddings of corresponding views -and using a score from fusing these aligned embeddings to select the next action to be taken. This is inspired by a dominant paradigm for modeling visual navigation tasks called viewpoint selection (Fried et al., 2018) where an agent predicts the next action by examining the resultant views each of those would produce and selecting the desired future view. Viewpoint selection is possible in some simulators such as R2R where the environment does not get altered by the agent's actions and the agent's movement is confined to a fixed grid. The ALFRED dataset uses the AI2-THOR simulator which supports a wider action space, physics modeling for movement and a more dynamic environment including irreversible actions. Hence, it is not possible to obtain the view that would result from an action without taking it, preventing direct application of viewpoint selection. Additionally, the agent must decide at each time step whether to perform navigation or manip-ulation actions. In contrast to prior work that uses a single classifier layer over all possible actions treating them equally, we propose a gate module which gives a higher weight to actions of a more relevant action type.",
"cite_spans": [
{
"start": 104,
"end": 125,
"text": "(Nguyen et al., 2021;",
"ref_id": "BIBREF11"
},
{
"start": 126,
"end": 143,
"text": "Kim et al., 2021)",
"ref_id": "BIBREF5"
},
{
"start": 663,
"end": 683,
"text": "(Fried et al., 2018)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We follow standard experimental procedure training our modified models on the train split and using success rate on the unseen validation split to compare to baselines and perform ablation studies. On this set, the proposed model equipped with the aforementioned modules outperforms the stateof-the-art multi-view setup approaches and the ablation study shows each proposed module helps improve the model's performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, we observe an unexpected and large performance gap between the unseen validation and test data splits. Our model outperforms stateof-the-art baseline models on the unseen validation split, but performs worse than them on the unseen test split. We hypothesize that it may be possible to overfit hyperparameters and design choices to one set of unseen environments (the unseen validation) and hence success on one such set of unseen environments is insufficient to guarantee that a model will generalize to another set of unseen environments (the unseen test). We report this finding as we believe this situation is likely more common during development on machine learning benchmarks, but such intermediate results are unlikely to be published. Instead after a poor result on a test set, it is likely that researchers continue further model modifications until a model setting is obtained that performs well on the test set. We believe that such models are likely overfitting to the test set of the benchmark and may not generalize well to a new test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we focus on improving models for the ALFRED (Shridhar et al., 2020) benchmark. ALFRED is built using the AI2-THOR simulator (Kolve et al., 2017) which consists of 120 indoor scenes across 4 types of rooms. Scenes also contain a diverse set of objects that are rearranged in different configurations for each trajectory in the dataset. In ALFRED, a agent is given a high level natural language goal statement (\"Put a chilled pan on the counter\") as well as step by step natural language instructions corresponding to subgoals to be completed in order for achieving the goal (\"Turn around and cross the room and then go right and turn to the left to face the stove ... Put the pan down on the counter to the right of the toaster\"). An agent has access to all these instructions at the start of the task and then has to iteratively predict navigation and manipulation actions in the environment based on egocentric image observations to complete subgoals in order. An agent must predict between a discrete set of possible navigation and manipulation actions, and predict a segmentation mask for the object to be manipulated if a manipulation action is predicted. The performance for an agent is evaluated by comparing the final states of the objects at the end of the action trajectory executed by the agent to the states of the objects at the end of the ground truth trajectory.",
"cite_spans": [
{
"start": 59,
"end": 82,
"text": "(Shridhar et al., 2020)",
"ref_id": "BIBREF16"
},
{
"start": 139,
"end": 159,
"text": "(Kolve et al., 2017)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset & Environment",
"sec_num": "2"
},
{
"text": "We employ a vision-language transformer, LXMERT (Tan and Bansal, 2019) as the base architecture for our model. We encode the language input using a learned word embedding and transformer layer, and action history using a linear layer. Following Pashevich et al. 2021, we extract image features using a faster R-CNN (Ren et al., 2015) pretrained on images from the AI2-THOR simulator, and average-pool features of regions into a single vector. The visual and action features are first combined via a liner layer, and then fused with language features through a cross modal transformer layer. View-Action Matching. We collect the multiple views (front, left, right, up, down) and go through the aforementioned process to obtain a feature V i from the cross modal transformer for each view, and compute its matching score M i with the corresponding action embedding A i using a feedforward network. Action-Type Gate. We additionally learn a gate vector using a linear layer over features of all views at the current time step to better distinguish between navigation and non-navigation actions. This layer is trained to predict high weights for actions of the same type as the ground truth action and low weights otherwise. The predicted weights are multiplied pointwise with match scores M i and the action with the highest resultant score is selected. For example, if the ground truth action at a particular time step is Move forward, the gate will ensure that a prediction of ToggleOff which is a non-navigation action will receive a higher loss than a prediction of Turn Right, which is also an incorrect action but of the same type as the ground truth action (navigation).",
"cite_spans": [
{
"start": 48,
"end": 70,
"text": "(Tan and Bansal, 2019)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3"
},
{
"text": "Loss. The model is trained via cross-entropy losses for action (teacher-forcing) and object type.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model",
"sec_num": "3"
},
{
"text": "Implementation & Training Details. We use 2 language and 2 cross-modal LXMERT layers for the model, and use 768 as the hidden size. We use AdamW (Loshchilov and Hutter, 2018) as the optimizer with the learning rate 1 \u00d7 10 \u22125 . All of the experiments are run on AWS 'p3.16xlarge' EC2 instances running Ubuntu 18.04. We employ PyTorch (Paszke et al., 2017) to build our models. Data Splits. Following Shridhar et al. 2020, we train our models on the train split and use success rate on the unseen validation split to perform model selection, and determine whether our model changes are likely to improve over existing state of the art models. We used the validation splits to evaluate the efficacy of variants of the transformer architecture, number of layers and number of epochs of training to use. We then submitted predictions from the best performing model on the unseen validation split to the evaluation server to obtain scores on the test sets. Evaluation Metrics. We report two evaluation metrics from Shridhar et al. (2020) on validation and test splits. Success rate (SR) measures the fraction of episodes whether the predicted model trajectory results in all object state changes produced by the ground truth action trajectory. Goal Condition Success Rate (GC) measures the fraction of such desired state changes across all episodes that were accomplished by model-predicted trajectories. Model Comparison. Recently, the best performing models on the ALFRED benchmark make use of semantic map representations of the environment (Blukis et al., 2021) . However, these rely on pre-exploration of the environment to build a semantic map, rather than utilizing language instruc- tions to directly navigate to target objects. Therefore, we focus on comparing our model with other multi-view setup models that are the state-of-theart among non-SLAM models. LWIT (Nguyen et al., 2021) predicts an initial actions from an selected instruction alone and integrates the actions sequence with visual information to generate final actions to take. ABP (Kim et al., 2021) factorizes the model into interactive perception and action policy modules for adapting to two different tasks (the former needs a pixel-level and the latter requires a global information). However, although they employ multi-view setup, the information from each view collapses into one integrated feature. On the other hand, our model exploit each view directly to keep the useful clues without any loss.",
"cite_spans": [
{
"start": 145,
"end": 174,
"text": "(Loshchilov and Hutter, 2018)",
"ref_id": "BIBREF10"
},
{
"start": 333,
"end": 354,
"text": "(Paszke et al., 2017)",
"ref_id": "BIBREF14"
},
{
"start": 1538,
"end": 1559,
"text": "(Blukis et al., 2021)",
"ref_id": "BIBREF3"
},
{
"start": 1861,
"end": 1887,
"text": "LWIT (Nguyen et al., 2021)",
"ref_id": null
},
{
"start": 2050,
"end": 2068,
"text": "(Kim et al., 2021)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "We first evaluate the utility of each modeling change on the unseen validation set of ALFRED. As shown Table 1 , we gain 4.6% on success rate from adding a wider field of view, an additional 2.5% from view-action matching and a further 2% from action type gating. We observe a variance of 3% in success rate of the same type of model trained with different random seeds so we consider a 4.6% improvement to be sufficiently large to be unlikely from pure variance.",
"cite_spans": [],
"ref_spans": [
{
"start": 103,
"end": 110,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Sub-Goal Performance. Considering the proportion of GotoLocation to the total number of sub-goal tasks (i.e., 48%) and its role of bridging other sub-goal tasks, navigation is very crucial ability for a agent to successfully perform this challenging ALFRED task. As shown in Table 2 , our full view-action matching (VAM) model improves the performance of GotoLocation task by 5.1% while also improves performance for some of other sub-goal tasks. This performance boost could attribute to the agent's ability to figure out where to go (View-Action Matching) and what to do (Action-Type Gate).",
"cite_spans": [],
"ref_spans": [
{
"start": 275,
"end": 282,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Validation-Test Performance Gap. When we compare to other baselines in Table 3 , although our model outperforms other state-of-the-art models on the unseen validation split by a large margin, its performance on the unseen test split is poorer, whereas the reverse trend is seen with ABP (Kim et al., 2021) . This suggests that good performance from a model on an unseen validation set may not be a good method to determine whether model changes are likely to generalize to another unseen test set.",
"cite_spans": [
{
"start": 287,
"end": 305,
"text": "(Kim et al., 2021)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 71,
"end": 78,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "This lack of generalization is more likely in current embodied learning tasks such as vision-andlanguage navigation or embodied task completion in comparison to other machine learning tasks due to the way unseen test sets are defined in embodied learning tasks. While ALFRED in particular does not introduce new object categories at test time, both validation and test unseen environments are visually different, by design from the training environment and from each other. When we compare models on the validation set, we hope that an increase in performance denotes a model that is more capable of generalizing to any unseen environment. However, it may only be the case that the model only generalizes better to the particular visual differences present in the unseen validation environment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "When the benchmark limits access to the test set, as in ALFRED, when dealing with a model that demonstrates variance when trained with different random seeds, hyperparameters and across training epochs, it is natural to choose the setting that results in the highest performance on the unseen validation set. However, a different setting may in fact be optimal for the unseen test set due to visual differences. While such a design is likely significantly more computationally expensive, it may be necessary to redesign benchmarks to take an average of performance from a few different variants of a model to reliably rank different modelling methods, instead of using scores from individual runs. We may also want to re-evaluate the value of keeping a test set private, as in the case of ALFRED that avoids prevents allowing models to overfit on the test set, but also makes it difficult to analyze the robustness of model performance between the validation and test sets. We would also like to encourage the reviewing community to enable the publication of modelling techniques whose performance is in the same ball- Table 3 : Success rate (%) on the ALFRED evaluation splits (GC: Goal-Condition). Our model outperforms the state-of-the-art multi-view setup models on validation splits but not test splits.",
"cite_spans": [],
"ref_spans": [
{
"start": 1119,
"end": 1126,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "park as existing state-of-the-art models, but novel in some way, as opposed to solely relying on a model achieving a top score on a leaderboard as a criterion for publication, as this limits the development that could be made using these alternative modeling approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "We attempted to improve a transformer model for embodied task completion by enabling it to effectively uses multiple views via view-action matching and action-type gating. Our view-action matching module computes a matching score between each a view and the embedding of the action used to generate it, and the gate module gives a higher weight to a more appropriate action type. While our model outperformed relevant baselines on the ALFRED unseen validation split, the trend was reversed on the unseen test split, suggesting that it may not be possible to over-utilize a validation split when making model selection choices so that the resultant model does not perform well on the test split. We choose to publish this result as we believe this phenomenon is likely more common than reported with machine learning benchmarks, but only noticeable to researchers when working on a benchmark with limited access to the test set. We additionally hope that our work encourages the publication of promising modelling approaches that do not work as reliably as expected, so that these can act as a guide to researchers to better inform their future directions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
}
],
"back_matter": [
{
"text": "We thank the reviewers for their helpful comments. This work was partially done while Hyounghun Kim was interning at Amazon Alexa AI and later extended at UNC, where it was supported by NSF Award 1840131 and DARPA KAIROS Grant FA8750-19-2-1004. The views contained in this article are those of the authors and not of the funding agency.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Vision-andlanguage navigation: Interpreting visually-grounded navigation instructions in real environments",
"authors": [
{
"first": "Peter",
"middle": [],
"last": "Anderson",
"suffix": ""
},
{
"first": "Qi",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Damien",
"middle": [],
"last": "Teney",
"suffix": ""
},
{
"first": "Jake",
"middle": [],
"last": "Bruce",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Johnson",
"suffix": ""
},
{
"first": "Niko",
"middle": [],
"last": "S\u00fcnderhauf",
"suffix": ""
},
{
"first": "Ian",
"middle": [],
"last": "Reid",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Gould",
"suffix": ""
},
{
"first": "Anton",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hengel",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
"volume": "",
"issue": "",
"pages": "3674--3683",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S\u00fcnderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. 2018. Vision-and- language navigation: Interpreting visually-grounded navigation instructions in real environments. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3674-3683.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Towards a dataset for human computer communication via grounded language acquisition",
"authors": [
{
"first": "Yonatan",
"middle": [],
"last": "Bisk",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
},
{
"first": "William",
"middle": [],
"last": "Wong",
"suffix": ""
}
],
"year": 2016,
"venue": "Workshops at the Thirtieth AAAI Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yonatan Bisk, Daniel Marcu, and William Wong. 2016. Towards a dataset for human computer communica- tion via grounded language acquisition. In Work- shops at the Thirtieth AAAI Conference on Artificial Intelligence.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Learning interpretable spatial operations in a rich 3d blocks world",
"authors": [
{
"first": "Yonatan",
"middle": [],
"last": "Bisk",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Kevin",
"suffix": ""
},
{
"first": "Yejin",
"middle": [],
"last": "Shih",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Choi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2018,
"venue": "Thirty-Second AAAI Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yonatan Bisk, Kevin J Shih, Yejin Choi, and Daniel Marcu. 2018. Learning interpretable spatial opera- tions in a rich 3d blocks world. In Thirty-Second AAAI Conference on Artificial Intelligence.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A persistent spatial semantic representation for high-level natural language instruction execution",
"authors": [
{
"first": "Valts",
"middle": [],
"last": "Blukis",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "Paxton",
"suffix": ""
},
{
"first": "Dieter",
"middle": [],
"last": "Fox",
"suffix": ""
},
{
"first": "Animesh",
"middle": [],
"last": "Garg",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Artzi",
"suffix": ""
}
],
"year": 2021,
"venue": "5th Annual Conference on Robot Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Valts Blukis, Chris Paxton, Dieter Fox, Animesh Garg, and Yoav Artzi. 2021. A persistent spatial semantic representation for high-level natural language instruc- tion execution. In 5th Annual Conference on Robot Learning.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Speaker-follower models for vision-and-language navigation",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Fried",
"suffix": ""
},
{
"first": "Ronghang",
"middle": [],
"last": "Hu",
"suffix": ""
},
{
"first": "Volkan",
"middle": [],
"last": "Cirik",
"suffix": ""
},
{
"first": "Anna",
"middle": [],
"last": "Rohrbach",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Andreas",
"suffix": ""
},
{
"first": "Louis-Philippe",
"middle": [],
"last": "Morency",
"suffix": ""
},
{
"first": "Taylor",
"middle": [],
"last": "Berg-Kirkpatrick",
"suffix": ""
},
{
"first": "Kate",
"middle": [],
"last": "Saenko",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Darrell",
"suffix": ""
}
],
"year": 2018,
"venue": "NeurIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. 2018. Speaker-follower models for vision-and-language navigation. In NeurIPS.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Agent with the big picture: Perceiving surroundings for interactive instruction following",
"authors": [
{
"first": "Byeonghwi",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Suvaansh",
"middle": [],
"last": "Bhambri",
"suffix": ""
},
{
"first": "Pratap",
"middle": [],
"last": "Singh",
"suffix": ""
},
{
"first": "Roozbeh",
"middle": [],
"last": "Mottaghi",
"suffix": ""
},
{
"first": "Jonghyun",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2021,
"venue": "Embodied AI Workshop CVPR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Byeonghwi Kim, Suvaansh Bhambri, Kunal Pratap Singh, Roozbeh Mottaghi, and Jonghyun Choi. 2021. Agent with the big picture: Perceiving surroundings for interactive instruction following. In Embodied AI Workshop CVPR.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Arramon: A joint navigation-assembly instruction interpretation task in dynamic environments",
"authors": [
{
"first": "Hyounghun",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Abhaysinh",
"middle": [],
"last": "Zala",
"suffix": ""
},
{
"first": "Graham",
"middle": [],
"last": "Burri",
"suffix": ""
},
{
"first": "Hao",
"middle": [],
"last": "Tan",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Bansal",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings",
"volume": "",
"issue": "",
"pages": "3910--3927",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hyounghun Kim, Abhaysinh Zala, Graham Burri, Hao Tan, and Mohit Bansal. 2020. Arramon: A joint navigation-assembly instruction interpretation task in dynamic environments. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 3910-3927.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv",
"authors": [
{
"first": "Eric",
"middle": [],
"last": "Kolve",
"suffix": ""
},
{
"first": "Roozbeh",
"middle": [],
"last": "Mottaghi",
"suffix": ""
},
{
"first": "Winson",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Eli",
"middle": [],
"last": "Van-Derbilt",
"suffix": ""
},
{
"first": "Luca",
"middle": [],
"last": "Weihs",
"suffix": ""
},
{
"first": "Alvaro",
"middle": [],
"last": "Herrasti",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Gordon",
"suffix": ""
},
{
"first": "Yuke",
"middle": [],
"last": "Zhu",
"suffix": ""
},
{
"first": "Abhinav",
"middle": [],
"last": "Gupta",
"suffix": ""
},
{
"first": "Ali",
"middle": [],
"last": "Farhadi",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli Van- derBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. 2017. AI2-THOR: An Interactive 3D Environment for Vi- sual AI. arXiv.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Ku",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Anderson",
"suffix": ""
},
{
"first": "Roma",
"middle": [],
"last": "Patel",
"suffix": ""
},
{
"first": "Eugene",
"middle": [],
"last": "Ie",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Baldridge",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "4392--4412",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. 2020. Room-across-room: Mul- tilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392-4412.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Spatial references and perspective in natural language instructions for collaborative manipulation",
"authors": [
{
"first": "Shen",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Rosario",
"middle": [],
"last": "Scalise",
"suffix": ""
},
{
"first": "Henny",
"middle": [],
"last": "Admoni",
"suffix": ""
},
{
"first": "Stephanie",
"middle": [],
"last": "Rosenthal",
"suffix": ""
},
{
"first": "Siddhartha",
"middle": [
"S"
],
"last": "Srinivasa",
"suffix": ""
}
],
"year": 2016,
"venue": "25th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN)",
"volume": "",
"issue": "",
"pages": "44--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shen Li, Rosario Scalise, Henny Admoni, Stephanie Rosenthal, and Siddhartha S Srinivasa. 2016. Spa- tial references and perspective in natural language instructions for collaborative manipulation. In 2016 25th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), pages 44-51. IEEE.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Decoupled weight decay regularization",
"authors": [
{
"first": "Ilya",
"middle": [],
"last": "Loshchilov",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Hutter",
"suffix": ""
}
],
"year": 2018,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Confer- ence on Learning Representations.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Look wide and interpret twice: Improving performance on interactive instructionfollowing tasks. IJCAI",
"authors": [
{
"first": "Masanori",
"middle": [],
"last": "Van-Quang Nguyen",
"suffix": ""
},
{
"first": "Takayuki",
"middle": [],
"last": "Suganuma",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Okatani",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. 2021. Look wide and interpret twice: Improving performance on interactive instruction- following tasks. IJCAI.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Gokhan Tur, and Dilek Hakkani-Tur. 2022. Teach: Task-driven embodied agents that chat",
"authors": [
{
"first": "Aishwarya",
"middle": [],
"last": "Padmakumar",
"suffix": ""
},
{
"first": "Jesse",
"middle": [],
"last": "Thomason",
"suffix": ""
},
{
"first": "Ayush",
"middle": [],
"last": "Shrivastava",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Lange",
"suffix": ""
},
{
"first": "Anjali",
"middle": [],
"last": "Narayan-Chen",
"suffix": ""
},
{
"first": "Spandana",
"middle": [],
"last": "Gella",
"suffix": ""
},
{
"first": "Robinson",
"middle": [],
"last": "Piramuthu",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aishwarya Padmakumar, Jesse Thomason, Ayush Shri- vastava, Patrick Lange, Anjali Narayan-Chen, Span- dana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. 2022. Teach: Task-driven em- bodied agents that chat. AAAI.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Episodic Transformer for Vision-and-Language Navigation",
"authors": [
{
"first": "Alexander",
"middle": [],
"last": "Pashevich",
"suffix": ""
},
{
"first": "Cordelia",
"middle": [],
"last": "Schmid",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2021,
"venue": "ICCV",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alexander Pashevich, Cordelia Schmid, and Chen Sun. 2021. Episodic Transformer for Vision-and- Language Navigation. In ICCV.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Automatic differentiation in PyTorch",
"authors": [
{
"first": "Adam",
"middle": [],
"last": "Paszke",
"suffix": ""
},
{
"first": "Sam",
"middle": [],
"last": "Gross",
"suffix": ""
},
{
"first": "Soumith",
"middle": [],
"last": "Chintala",
"suffix": ""
},
{
"first": "Gregory",
"middle": [],
"last": "Chanan",
"suffix": ""
},
{
"first": "Edward",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zachary",
"middle": [],
"last": "Devito",
"suffix": ""
},
{
"first": "Zeming",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Alban",
"middle": [],
"last": "Desmaison",
"suffix": ""
},
{
"first": "Luca",
"middle": [],
"last": "Antiga",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Lerer",
"suffix": ""
}
],
"year": 2017,
"venue": "NIPS Autodiff Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in PyTorch. In NIPS Autodiff Workshop.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Faster r-cnn: Towards real-time object detection with region proposal networks",
"authors": [
{
"first": "Kaiming",
"middle": [],
"last": "Shaoqing Ren",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Ross",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Girshick",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2015,
"venue": "NIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shaoqing Ren, Kaiming He, Ross B Girshick, and Jian Sun. 2015. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks",
"authors": [
{
"first": "Mohit",
"middle": [],
"last": "Shridhar",
"suffix": ""
},
{
"first": "Jesse",
"middle": [],
"last": "Thomason",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Gordon",
"suffix": ""
},
{
"first": "Yonatan",
"middle": [],
"last": "Bisk",
"suffix": ""
},
{
"first": "Winson",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Roozbeh",
"middle": [],
"last": "Mottaghi",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "Dieter",
"middle": [],
"last": "Fox",
"suffix": ""
}
],
"year": 2020,
"venue": "The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. 2020. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR).",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Lxmert: Learning cross-modality encoder representations from transformers",
"authors": [
{
"first": "Hao",
"middle": [],
"last": "Tan",
"suffix": ""
},
{
"first": "Mohit",
"middle": [],
"last": "Bansal",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hao Tan and Mohit Bansal. 2019. Lxmert: Learning cross-modality encoder representations from trans- formers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Vision-and-dialog navigation",
"authors": [
{
"first": "Jesse",
"middle": [],
"last": "Thomason",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Murray",
"suffix": ""
},
{
"first": "Maya",
"middle": [],
"last": "Cakmak",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2019,
"venue": "Conference on Robot Learning (CoRL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. 2019. Vision-and-dialog naviga- tion. In Conference on Robot Learning (CoRL).",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Learning language games through interaction",
"authors": [
{
"first": "I",
"middle": [],
"last": "Sida",
"suffix": ""
},
{
"first": "Percy",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Christopher D",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "2368--2378",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sida I Wang, Percy Liang, and Christopher D Manning. 2016. Learning language games through interaction. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2368-2378.",
"links": null
}
},
"ref_entries": {
"TABREF2": {
"content": "<table/>",
"type_str": "table",
"html": null,
"text": "Success rate (%) of the sub-goal tasks on the ALFRED validation unseen split.",
"num": null
},
"TABREF3": {
"content": "<table><tr><td>Split</td><td>Model</td><td>SR</td><td colspan=\"2\">Seen</td><td>GC</td><td>Unseen SR GC</td></tr><tr><td/><td>LWIT</td><td colspan=\"5\">33.70 43.10 9.70 23.10</td></tr><tr><td>Val</td><td>ABP</td><td colspan=\"5\">42.93 50.45 12.55 25.19</td></tr><tr><td/><td colspan=\"3\">VAM (Ours) 40.9</td><td colspan=\"2\">47.9</td><td>13.8</td><td>28.1</td></tr></table>",
"type_str": "table",
"html": null,
"text": "Test LWIT 29.16 38.82 8.37 19.13 ABP 44.55 51.13 15.43 24.76 VAM (Ours) 35.42 43.98 8.57 20.69",
"num": null
}
}
}
} |