{ "streams": [ [ "view", "camera view" ], [ "tl", "tactile left" ], [ "tr", "tactile right" ] ], "metrics": [ [ "psnr", "PSNR (dB)", 1 ], [ "ssim", "SSIM", 1 ], [ "lpips", "LPIPS", -1 ] ], "arms": [ { "key": "delta", "label": "delta_action + isolate", "job": "514906", "ckpt": "best-epoch=38-val_loss=0.0901.ckpt", "note": "per-step rows (force f[k+1]-f[k]); gels do not attend each other", "wandb": { "epoch": 99, "val_loss": 0.09502935409545898, "val_loss_video": 0.06267819553613663, "val_loss_tactile": 0.032351139932870865, "train_loss_epoch": 0.06542341411113739 }, "best_val": 0.0901 }, { "key": "chunk", "label": "action_chunking + isolate", "job": "514907", "ckpt": "best-epoch=33-val_loss=0.0902.ckpt", "note": "anchor-relative rows (force f[k+1]-f[0]); gels do not attend each other", "wandb": { "epoch": 99, "val_loss": 0.09490643441677094, "val_loss_video": 0.06269224733114243, "val_loss_tactile": 0.032214194536209106, "train_loss_epoch": 0.0656326413154602 }, "best_val": 0.0902 }, { "key": "noisolate", "label": "delta_action, no isolate", "job": "516058", "ckpt": "best-epoch=33-val_loss=0.0901.ckpt", "note": "same as 514906 with isolate_sensors: false -- the gels attend each other", "wandb": { "epoch": 99, "val_loss": 0.09476783871650696, "val_loss_video": 0.06237725913524628, "val_loss_tactile": 0.03239057958126068, "train_loss_epoch": 0.06542208790779114 }, "best_val": 0.0901 } ], "cells": [ { "label": "delta_action + isolate (job 514906)", "n": 24, "hl": true, "key": "delta", "m": { "view_psnr": 24.302958333333333, "view_ssim": 0.8259875000000001, "view_lpips": 0.06796125, "tl_psnr": 44.4156375, "tl_ssim": 0.9813075000000001, "tl_lpips": 0.03966166666666667, "tr_psnr": 41.47065833333334, "tr_ssim": 0.9762179166666667, "tr_lpips": 0.06738791666666667 }, "win": { "view_psnr": true, "view_ssim": true, "view_lpips": true, "tl_psnr": false, "tl_ssim": false, "tl_lpips": true, "tr_psnr": false, "tr_ssim": false, "tr_lpips": false } }, { "label": "action_chunking + isolate (job 514907)", "n": 24, "hl": true, "key": "chunk", "m": { "view_psnr": 23.993458333333333, "view_ssim": 0.8206920833333333, "view_lpips": 0.07317708333333334, "tl_psnr": 44.47287500000001, "tl_ssim": 0.9821879166666667, "tl_lpips": 0.041695, "tr_psnr": 42.23287916666667, "tr_ssim": 0.97830625, "tr_lpips": 0.06903416666666666 }, "win": { "view_psnr": false, "view_ssim": false, "view_lpips": false, "tl_psnr": false, "tl_ssim": false, "tl_lpips": false, "tr_psnr": false, "tr_ssim": false, "tr_lpips": false } }, { "label": "delta_action, no isolate (job 516058)", "n": 24, "hl": true, "key": "noisolate", "m": { "view_psnr": 24.023112499999996, "view_ssim": 0.8225966666666666, "view_lpips": 0.07142, "tl_psnr": 44.736066666666666, "tl_ssim": 0.9828899999999999, "tl_lpips": 0.04015041666666667, "tr_psnr": 42.302770833333334, "tr_ssim": 0.9789804166666668, "tr_lpips": 0.06434333333333334 }, "win": { "view_psnr": false, "view_ssim": false, "view_lpips": false, "tl_psnr": true, "tl_ssim": true, "tl_lpips": false, "tr_psnr": true, "tr_ssim": true, "tr_lpips": true } } ], "abs_ref": [ { "label": "a9dfilm_fps15 (absolute force)", "n": 10, "hl": false, "m": { "view_psnr": 23.8438, "view_ssim": 0.819633, "view_lpips": 0.07308700000000001, "tl_psnr": 44.28192, "tl_ssim": 0.9809570000000001, "tl_lpips": 0.04113100000000001, "tr_psnr": 43.48711000000001, "tr_ssim": 0.9820199999999998, "tr_lpips": 0.059044 } }, { "label": "a9dfilm_isolate_fps15 (absolute force)", "n": 10, "hl": false, "m": { "view_psnr": 24.3308, "view_ssim": 0.8274779999999999, "view_lpips": 0.069304, "tl_psnr": 43.90608999999999, "tl_ssim": 0.979203, "tl_lpips": 0.039718, "tr_psnr": 42.00175, "tr_ssim": 0.979185, "tr_lpips": 0.06392499999999998 } } ], "sign": [ { "title": "delta_action vs action_chunking (the action representation)", "line": "data.action_repr", "a": "delta_action + isolate", "b": "action_chunking + isolate", "rows": [ { "name": "camera view — PSNR (dB)", "a": 24.302958333333333, "b": 23.993458333333336, "median": 0.3224, "wins": 18, "n": 24, "p": 0.022655844688415527, "nd": 2 }, { "name": "camera view — SSIM", "a": 0.8259875000000002, "b": 0.8206920833333333, "median": 0.005294999999999939, "wins": 19, "n": 24, "p": 0.006610751152038574, "nd": 4 }, { "name": "camera view — LPIPS", "a": 0.06796125, "b": 0.07317708333333332, "median": -0.0061200000000000004, "wins": 20, "n": 24, "p": 0.001543879508972168, "nd": 4 }, { "name": "tactile left — PSNR (dB)", "a": 44.4156375, "b": 44.47287500000001, "median": -0.1045499999999997, "wins": 7, "n": 24, "p": 0.06391465663909912, "nd": 2 }, { "name": "tactile left — SSIM", "a": 0.9813075000000001, "b": 0.9821879166666666, "median": -0.000140000000000029, "wins": 8, "n": 24, "p": 0.15158963203430176, "nd": 4 }, { "name": "tactile left — LPIPS", "a": 0.03966166666666667, "b": 0.041695, "median": -6.999999999999932e-05, "wins": 14, "n": 24, "p": 0.5412561893463135, "nd": 4 }, { "name": "tactile right — PSNR (dB)", "a": 41.47065833333333, "b": 42.23287916666667, "median": -0.48385000000000034, "wins": 5, "n": 24, "p": 0.006610751152038574, "nd": 2 }, { "name": "tactile right — SSIM", "a": 0.9762179166666668, "b": 0.9783062499999998, "median": -0.0013649999999999496, "wins": 5, "n": 24, "p": 0.006610751152038574, "nd": 4 }, { "name": "tactile right — LPIPS", "a": 0.06738791666666666, "b": 0.06903416666666669, "median": 0.002065000000000001, "wins": 9, "n": 24, "p": 0.30745625495910645, "nd": 4 } ] }, { "title": "isolate_sensors true vs false (do the two gels attend each other?)", "line": "model.isolate_sensors", "a": "delta_action + isolate", "b": "delta_action, no isolate", "rows": [ { "name": "camera view — PSNR (dB)", "a": 24.302958333333333, "b": 24.0231125, "median": 0.26039999999999885, "wins": 15, "n": 24, "p": 0.30745625495910645, "nd": 2 }, { "name": "camera view — SSIM", "a": 0.8259875000000002, "b": 0.8225966666666668, "median": 0.0031200000000000117, "wins": 15, "n": 24, "p": 0.30745625495910645, "nd": 4 }, { "name": "camera view — LPIPS", "a": 0.06796125, "b": 0.07142, "median": -0.00213, "wins": 16, "n": 24, "p": 0.15158963203430176, "nd": 4 }, { "name": "tactile left — PSNR (dB)", "a": 44.4156375, "b": 44.736066666666666, "median": -0.32005000000000194, "wins": 7, "n": 24, "p": 0.06391465663909912, "nd": 2 }, { "name": "tactile left — SSIM", "a": 0.9813075000000001, "b": 0.9828899999999998, "median": -0.00023499999999998522, "wins": 4, "n": 24, "p": 0.001543879508972168, "nd": 4 }, { "name": "tactile left — LPIPS", "a": 0.03966166666666667, "b": 0.04015041666666667, "median": 0.00011000000000000202, "wins": 8, "n": 24, "p": 0.15158963203430176, "nd": 4 }, { "name": "tactile right — PSNR (dB)", "a": 41.47065833333333, "b": 42.302770833333334, "median": -0.8111499999999978, "wins": 8, "n": 24, "p": 0.15158963203430176, "nd": 2 }, { "name": "tactile right — SSIM", "a": 0.9762179166666668, "b": 0.9789804166666666, "median": -0.0016900000000000248, "wins": 5, "n": 24, "p": 0.006610751152038574, "nd": 4 }, { "name": "tactile right — LPIPS", "a": 0.06738791666666666, "b": 0.06434333333333335, "median": 0.000915, "wins": 9, "n": 24, "p": 0.30745625495910645, "nd": 4 } ] } ], "samples": [ { "sample": "sample_000", "episode": "motherboard_0510_episode_006.pt", "t_start": 1210, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.0637, "ssim": 0.82581, "lpips": 0.07462 }, "tl": { "psnr": 38.8116, "ssim": 0.95857, "lpips": 0.07213 }, "tr": { "psnr": 43.4656, "ssim": 0.99051, "lpips": 0.02442 } }, "chunk": { "view": { "psnr": 23.9987, "ssim": 0.82522, "lpips": 0.07053 }, "tl": { "psnr": 38.4499, "ssim": 0.95984, "lpips": 0.09006 }, "tr": { "psnr": 45.8582, "ssim": 0.99147, "lpips": 0.02104 } }, "noisolate": { "view": { "psnr": 23.7172, "ssim": 0.81939, "lpips": 0.074 }, "tl": { "psnr": 38.3388, "ssim": 0.96087, "lpips": 0.08885 }, "tr": { "psnr": 46.6219, "ssim": 0.99224, "lpips": 0.01836 } } } }, { "sample": "sample_001", "episode": "motherboard_0510_episode_006.pt", "t_start": 1355, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.2419, "ssim": 0.80027, "lpips": 0.08689 }, "tl": { "psnr": 41.6415, "ssim": 0.97767, "lpips": 0.0314 }, "tr": { "psnr": 43.5466, "ssim": 0.98362, "lpips": 0.06314 } }, "chunk": { "view": { "psnr": 22.9977, "ssim": 0.79729, "lpips": 0.09214 }, "tl": { "psnr": 41.7217, "ssim": 0.97724, "lpips": 0.03191 }, "tr": { "psnr": 45.5447, "ssim": 0.98734, "lpips": 0.05732 } }, "noisolate": { "view": { "psnr": 23.1239, "ssim": 0.80347, "lpips": 0.0861 }, "tl": { "psnr": 41.1616, "ssim": 0.97697, "lpips": 0.03509 }, "tr": { "psnr": 46.6428, "ssim": 0.98982, "lpips": 0.04561 } } } }, { "sample": "sample_002", "episode": "motherboard_0510_episode_005.pt", "t_start": 165, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.9418, "ssim": 0.83672, "lpips": 0.06885 }, "tl": { "psnr": 49.4592, "ssim": 0.99311, "lpips": 0.00695 }, "tr": { "psnr": 40.3886, "ssim": 0.97903, "lpips": 0.05849 } }, "chunk": { "view": { "psnr": 23.8833, "ssim": 0.82202, "lpips": 0.0857 }, "tl": { "psnr": 50.054, "ssim": 0.99339, "lpips": 0.00681 }, "tr": { "psnr": 40.1525, "ssim": 0.97798, "lpips": 0.05624 } }, "noisolate": { "view": { "psnr": 23.9451, "ssim": 0.82664, "lpips": 0.08634 }, "tl": { "psnr": 50.243, "ssim": 0.99342, "lpips": 0.00679 }, "tr": { "psnr": 40.0383, "ssim": 0.97836, "lpips": 0.05882 } } } }, { "sample": "sample_003", "episode": "motherboard_0510_episode_006.pt", "t_start": 693, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 26.3211, "ssim": 0.86865, "lpips": 0.05631 }, "tl": { "psnr": 44.3058, "ssim": 0.98546, "lpips": 0.02207 }, "tr": { "psnr": 40.2477, "ssim": 0.96148, "lpips": 0.07096 } }, "chunk": { "view": { "psnr": 26.4132, "ssim": 0.87175, "lpips": 0.05534 }, "tl": { "psnr": 43.051, "ssim": 0.9852, "lpips": 0.03153 }, "tr": { "psnr": 39.7445, "ssim": 0.95902, "lpips": 0.08888 } }, "noisolate": { "view": { "psnr": 26.3662, "ssim": 0.86887, "lpips": 0.05451 }, "tl": { "psnr": 44.0982, "ssim": 0.98557, "lpips": 0.02392 }, "tr": { "psnr": 41.2067, "ssim": 0.968, "lpips": 0.07652 } } } }, { "sample": "sample_004", "episode": "motherboard_0510_episode_006.pt", "t_start": 1727, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.6689, "ssim": 0.82315, "lpips": 0.06748 }, "tl": { "psnr": 49.4617, "ssim": 0.99304, "lpips": 0.0077 }, "tr": { "psnr": 42.4214, "ssim": 0.98417, "lpips": 0.03772 } }, "chunk": { "view": { "psnr": 23.2447, "ssim": 0.8129, "lpips": 0.07543 }, "tl": { "psnr": 49.5984, "ssim": 0.99322, "lpips": 0.00783 }, "tr": { "psnr": 43.3189, "ssim": 0.98478, "lpips": 0.03814 } }, "noisolate": { "view": { "psnr": 23.7351, "ssim": 0.82303, "lpips": 0.06958 }, "tl": { "psnr": 50.0605, "ssim": 0.99314, "lpips": 0.00749 }, "tr": { "psnr": 42.1065, "ssim": 0.98359, "lpips": 0.0454 } } } }, { "sample": "sample_005", "episode": "motherboard_0510_episode_006.pt", "t_start": 1623, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 22.79, "ssim": 0.79491, "lpips": 0.09064 }, "tl": { "psnr": 38.5151, "ssim": 0.95189, "lpips": 0.0677 }, "tr": { "psnr": 44.6409, "ssim": 0.98911, "lpips": 0.03296 } }, "chunk": { "view": { "psnr": 23.4135, "ssim": 0.8098, "lpips": 0.07623 }, "tl": { "psnr": 39.3184, "ssim": 0.95984, "lpips": 0.08279 }, "tr": { "psnr": 48.847, "ssim": 0.99237, "lpips": 0.01276 } }, "noisolate": { "view": { "psnr": 22.8831, "ssim": 0.80133, "lpips": 0.08125 }, "tl": { "psnr": 39.3103, "ssim": 0.96174, "lpips": 0.09149 }, "tr": { "psnr": 46.4857, "ssim": 0.99104, "lpips": 0.01926 } } } }, { "sample": "sample_006", "episode": "motherboard_0510_episode_006.pt", "t_start": 1291, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.1065, "ssim": 0.82112, "lpips": 0.07034 }, "tl": { "psnr": 38.8002, "ssim": 0.96871, "lpips": 0.07511 }, "tr": { "psnr": 42.6866, "ssim": 0.9821, "lpips": 0.12858 } }, "chunk": { "view": { "psnr": 23.3685, "ssim": 0.80685, "lpips": 0.08058 }, "tl": { "psnr": 39.516, "ssim": 0.97021, "lpips": 0.06521 }, "tr": { "psnr": 42.6964, "ssim": 0.98179, "lpips": 0.12849 } }, "noisolate": { "view": { "psnr": 23.1528, "ssim": 0.7995, "lpips": 0.0813 }, "tl": { "psnr": 41.0983, "ssim": 0.97576, "lpips": 0.04426 }, "tr": { "psnr": 42.7962, "ssim": 0.98218, "lpips": 0.13296 } } } }, { "sample": "sample_007", "episode": "motherboard_0510_episode_006.pt", "t_start": 875, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 25.3625, "ssim": 0.84214, "lpips": 0.05222 }, "tl": { "psnr": 39.1007, "ssim": 0.97569, "lpips": 0.09586 }, "tr": { "psnr": 46.5665, "ssim": 0.99165, "lpips": 0.01048 } }, "chunk": { "view": { "psnr": 24.4187, "ssim": 0.8233, "lpips": 0.05935 }, "tl": { "psnr": 39.2773, "ssim": 0.97582, "lpips": 0.09385 }, "tr": { "psnr": 47.0366, "ssim": 0.99187, "lpips": 0.00932 } }, "noisolate": { "view": { "psnr": 24.5308, "ssim": 0.83139, "lpips": 0.05709 }, "tl": { "psnr": 39.2677, "ssim": 0.97565, "lpips": 0.0958 }, "tr": { "psnr": 46.528, "ssim": 0.99181, "lpips": 0.00942 } } } }, { "sample": "sample_008", "episode": "motherboard_0510_episode_006.pt", "t_start": 1585, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.6055, "ssim": 0.81381, "lpips": 0.06466 }, "tl": { "psnr": 49.489, "ssim": 0.99289, "lpips": 0.00772 }, "tr": { "psnr": 36.0776, "ssim": 0.96487, "lpips": 0.07809 } }, "chunk": { "view": { "psnr": 23.1955, "ssim": 0.80699, "lpips": 0.07089 }, "tl": { "psnr": 49.6179, "ssim": 0.99304, "lpips": 0.00783 }, "tr": { "psnr": 38.0228, "ssim": 0.97082, "lpips": 0.08625 } }, "noisolate": { "view": { "psnr": 23.3673, "ssim": 0.81184, "lpips": 0.06741 }, "tl": { "psnr": 49.4626, "ssim": 0.99306, "lpips": 0.00768 }, "tr": { "psnr": 37.2409, "ssim": 0.96884, "lpips": 0.0973 } } } }, { "sample": "sample_009", "episode": "motherboard_0510_episode_006.pt", "t_start": 1099, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.4834, "ssim": 0.83061, "lpips": 0.07256 }, "tl": { "psnr": 49.5255, "ssim": 0.99297, "lpips": 0.00803 }, "tr": { "psnr": 38.0824, "ssim": 0.95946, "lpips": 0.13977 } }, "chunk": { "view": { "psnr": 23.5276, "ssim": 0.80833, "lpips": 0.08093 }, "tl": { "psnr": 49.6981, "ssim": 0.99312, "lpips": 0.00806 }, "tr": { "psnr": 38.9258, "ssim": 0.96412, "lpips": 0.12928 } }, "noisolate": { "view": { "psnr": 24.0039, "ssim": 0.81948, "lpips": 0.07472 }, "tl": { "psnr": 49.9435, "ssim": 0.99322, "lpips": 0.00801 }, "tr": { "psnr": 39.8641, "ssim": 0.97077, "lpips": 0.11137 } } } }, { "sample": "sample_010", "episode": "motherboard_0510_episode_006.pt", "t_start": 527, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 26.5443, "ssim": 0.87445, "lpips": 0.04651 }, "tl": { "psnr": 44.7002, "ssim": 0.98802, "lpips": 0.03601 }, "tr": { "psnr": 40.5382, "ssim": 0.96482, "lpips": 0.04882 } }, "chunk": { "view": { "psnr": 26.3784, "ssim": 0.86687, "lpips": 0.04627 }, "tl": { "psnr": 45.0662, "ssim": 0.98854, "lpips": 0.03653 }, "tr": { "psnr": 40.8327, "ssim": 0.97473, "lpips": 0.12449 } }, "noisolate": { "view": { "psnr": 25.6938, "ssim": 0.85597, "lpips": 0.04938 }, "tl": { "psnr": 45.4799, "ssim": 0.98993, "lpips": 0.03115 }, "tr": { "psnr": 41.2338, "ssim": 0.97489, "lpips": 0.09489 } } } }, { "sample": "sample_011", "episode": "motherboard_0510_episode_006.pt", "t_start": 1700, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.1103, "ssim": 0.79049, "lpips": 0.09486 }, "tl": { "psnr": 49.0372, "ssim": 0.99228, "lpips": 0.00814 }, "tr": { "psnr": 39.2186, "ssim": 0.97118, "lpips": 0.05607 } }, "chunk": { "view": { "psnr": 23.4052, "ssim": 0.79529, "lpips": 0.10072 }, "tl": { "psnr": 48.9773, "ssim": 0.9923, "lpips": 0.00846 }, "tr": { "psnr": 38.2752, "ssim": 0.96725, "lpips": 0.07909 } }, "noisolate": { "view": { "psnr": 24.3316, "ssim": 0.81192, "lpips": 0.08308 }, "tl": { "psnr": 49.3631, "ssim": 0.99247, "lpips": 0.00798 }, "tr": { "psnr": 40.1558, "ssim": 0.97513, "lpips": 0.05593 } } } }, { "sample": "sample_012", "episode": "motherboard_0510_episode_006.pt", "t_start": 203, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 25.8985, "ssim": 0.85631, "lpips": 0.0521 }, "tl": { "psnr": 41.2565, "ssim": 0.98116, "lpips": 0.04532 }, "tr": { "psnr": 39.289, "ssim": 0.96742, "lpips": 0.08314 } }, "chunk": { "view": { "psnr": 25.5209, "ssim": 0.85542, "lpips": 0.05437 }, "tl": { "psnr": 41.0633, "ssim": 0.98047, "lpips": 0.04519 }, "tr": { "psnr": 40.5636, "ssim": 0.97099, "lpips": 0.06737 } }, "noisolate": { "view": { "psnr": 25.5027, "ssim": 0.8514, "lpips": 0.05854 }, "tl": { "psnr": 41.3981, "ssim": 0.98154, "lpips": 0.04334 }, "tr": { "psnr": 39.1574, "ssim": 0.96838, "lpips": 0.07721 } } } }, { "sample": "sample_013", "episode": "motherboard_0510_episode_006.pt", "t_start": 787, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.9867, "ssim": 0.84019, "lpips": 0.06595 }, "tl": { "psnr": 45.5469, "ssim": 0.98809, "lpips": 0.02109 }, "tr": { "psnr": 42.67, "ssim": 0.98426, "lpips": 0.05839 } }, "chunk": { "view": { "psnr": 25.0387, "ssim": 0.8363, "lpips": 0.06718 }, "tl": { "psnr": 44.9957, "ssim": 0.98623, "lpips": 0.01972 }, "tr": { "psnr": 45.3747, "ssim": 0.98778, "lpips": 0.03962 } }, "noisolate": { "view": { "psnr": 25.0131, "ssim": 0.83891, "lpips": 0.06495 }, "tl": { "psnr": 44.6728, "ssim": 0.98705, "lpips": 0.0294 }, "tr": { "psnr": 43.7227, "ssim": 0.9858, "lpips": 0.05767 } } } }, { "sample": "sample_014", "episode": "motherboard_0510_episode_006.pt", "t_start": 205, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.658, "ssim": 0.83622, "lpips": 0.05604 }, "tl": { "psnr": 39.0593, "ssim": 0.96964, "lpips": 0.07429 }, "tr": { "psnr": 41.8324, "ssim": 0.97796, "lpips": 0.06172 } }, "chunk": { "view": { "psnr": 25.0625, "ssim": 0.8466, "lpips": 0.05914 }, "tl": { "psnr": 38.3939, "ssim": 0.97014, "lpips": 0.09329 }, "tr": { "psnr": 42.6104, "ssim": 0.97891, "lpips": 0.048 } }, "noisolate": { "view": { "psnr": 24.8783, "ssim": 0.84515, "lpips": 0.05673 }, "tl": { "psnr": 38.9854, "ssim": 0.97225, "lpips": 0.08102 }, "tr": { "psnr": 43.1013, "ssim": 0.98095, "lpips": 0.02887 } } } }, { "sample": "sample_015", "episode": "motherboard_0510_episode_006.pt", "t_start": 21, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.493, "ssim": 0.8224, "lpips": 0.06877 }, "tl": { "psnr": 49.5656, "ssim": 0.99357, "lpips": 0.00878 }, "tr": { "psnr": 36.8908, "ssim": 0.96867, "lpips": 0.15244 } }, "chunk": { "view": { "psnr": 22.6669, "ssim": 0.80601, "lpips": 0.0789 }, "tl": { "psnr": 49.5932, "ssim": 0.99353, "lpips": 0.00838 }, "tr": { "psnr": 36.9777, "ssim": 0.97044, "lpips": 0.1352 } }, "noisolate": { "view": { "psnr": 23.5834, "ssim": 0.82599, "lpips": 0.07051 }, "tl": { "psnr": 50.4054, "ssim": 0.99379, "lpips": 0.00818 }, "tr": { "psnr": 36.6049, "ssim": 0.96871, "lpips": 0.14968 } } } }, { "sample": "sample_016", "episode": "motherboard_0510_episode_006.pt", "t_start": 659, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.8871, "ssim": 0.81558, "lpips": 0.06191 }, "tl": { "psnr": 39.2804, "ssim": 0.96892, "lpips": 0.08009 }, "tr": { "psnr": 43.6734, "ssim": 0.98553, "lpips": 0.04452 } }, "chunk": { "view": { "psnr": 23.3793, "ssim": 0.80888, "lpips": 0.06757 }, "tl": { "psnr": 39.5328, "ssim": 0.97249, "lpips": 0.08214 }, "tr": { "psnr": 42.7353, "ssim": 0.98328, "lpips": 0.04644 } }, "noisolate": { "view": { "psnr": 23.2486, "ssim": 0.80857, "lpips": 0.06921 }, "tl": { "psnr": 39.7129, "ssim": 0.97299, "lpips": 0.0758 }, "tr": { "psnr": 43.3122, "ssim": 0.98403, "lpips": 0.04481 } } } }, { "sample": "sample_017", "episode": "motherboard_0510_episode_006.pt", "t_start": 1814, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.6111, "ssim": 0.82247, "lpips": 0.06471 }, "tl": { "psnr": 49.6827, "ssim": 0.9933, "lpips": 0.00684 }, "tr": { "psnr": 43.512, "ssim": 0.98328, "lpips": 0.03755 } }, "chunk": { "view": { "psnr": 23.2374, "ssim": 0.81436, "lpips": 0.07532 }, "tl": { "psnr": 50.0857, "ssim": 0.99339, "lpips": 0.0068 }, "tr": { "psnr": 43.7571, "ssim": 0.98409, "lpips": 0.03457 } }, "noisolate": { "view": { "psnr": 22.9016, "ssim": 0.81314, "lpips": 0.07866 }, "tl": { "psnr": 49.9969, "ssim": 0.99333, "lpips": 0.00664 }, "tr": { "psnr": 42.8842, "ssim": 0.98262, "lpips": 0.04763 } } } }, { "sample": "sample_018", "episode": "motherboard_0510_episode_006.pt", "t_start": 234, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.3544, "ssim": 0.80581, "lpips": 0.08201 }, "tl": { "psnr": 34.471, "ssim": 0.94292, "lpips": 0.17849 }, "tr": { "psnr": 50.8226, "ssim": 0.99481, "lpips": 0.00375 } }, "chunk": { "view": { "psnr": 23.011, "ssim": 0.79745, "lpips": 0.09042 }, "tl": { "psnr": 34.879, "ssim": 0.94962, "lpips": 0.17845 }, "tr": { "psnr": 51.101, "ssim": 0.99487, "lpips": 0.00398 } }, "noisolate": { "view": { "psnr": 23.0718, "ssim": 0.80154, "lpips": 0.08375 }, "tl": { "psnr": 34.9839, "ssim": 0.94962, "lpips": 0.17699 }, "tr": { "psnr": 51.2128, "ssim": 0.99493, "lpips": 0.00382 } } } }, { "sample": "sample_019", "episode": "motherboard_0510_episode_006.pt", "t_start": 903, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.5727, "ssim": 0.83193, "lpips": 0.05267 }, "tl": { "psnr": 50.1773, "ssim": 0.9934, "lpips": 0.00665 }, "tr": { "psnr": 44.2344, "ssim": 0.98712, "lpips": 0.02677 } }, "chunk": { "view": { "psnr": 24.3546, "ssim": 0.82907, "lpips": 0.05935 }, "tl": { "psnr": 50.2374, "ssim": 0.99345, "lpips": 0.00692 }, "tr": { "psnr": 44.3485, "ssim": 0.98752, "lpips": 0.02489 } }, "noisolate": { "view": { "psnr": 24.4828, "ssim": 0.83227, "lpips": 0.05811 }, "tl": { "psnr": 50.1359, "ssim": 0.99337, "lpips": 0.00702 }, "tr": { "psnr": 47.2453, "ssim": 0.99128, "lpips": 0.01198 } } } }, { "sample": "sample_020", "episode": "motherboard_0510_episode_006.pt", "t_start": 37, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.991, "ssim": 0.83426, "lpips": 0.07373 }, "tl": { "psnr": 37.2768, "ssim": 0.97357, "lpips": 0.06717 }, "tr": { "psnr": 43.4487, "ssim": 0.9842, "lpips": 0.0379 } }, "chunk": { "view": { "psnr": 24.1049, "ssim": 0.83448, "lpips": 0.07974 }, "tl": { "psnr": 37.3567, "ssim": 0.9753, "lpips": 0.06425 }, "tr": { "psnr": 44.4416, "ssim": 0.98664, "lpips": 0.02987 } }, "noisolate": { "view": { "psnr": 24.1544, "ssim": 0.83555, "lpips": 0.07291 }, "tl": { "psnr": 37.3996, "ssim": 0.97555, "lpips": 0.06362 }, "tr": { "psnr": 44.0421, "ssim": 0.98585, "lpips": 0.03713 } } } }, { "sample": "sample_021", "episode": "motherboard_0510_episode_005.pt", "t_start": 302, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 25.1474, "ssim": 0.83792, "lpips": 0.05935 }, "tl": { "psnr": 47.5451, "ssim": 0.99068, "lpips": 0.00932 }, "tr": { "psnr": 37.891, "ssim": 0.9616, "lpips": 0.15989 } }, "chunk": { "view": { "psnr": 24.1582, "ssim": 0.82379, "lpips": 0.07024 }, "tl": { "psnr": 47.7784, "ssim": 0.99047, "lpips": 0.00967 }, "tr": { "psnr": 37.4835, "ssim": 0.96876, "lpips": 0.181 } }, "noisolate": { "view": { "psnr": 25.3278, "ssim": 0.84441, "lpips": 0.0591 }, "tl": { "psnr": 48.4463, "ssim": 0.99205, "lpips": 0.00804 }, "tr": { "psnr": 39.0386, "ssim": 0.97135, "lpips": 0.12352 } } } }, { "sample": "sample_022", "episode": "motherboard_0510_episode_006.pt", "t_start": 985, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 24.3537, "ssim": 0.8169, "lpips": 0.06912 }, "tl": { "psnr": 49.9585, "ssim": 0.9933, "lpips": 0.00727 }, "tr": { "psnr": 39.2635, "ssim": 0.96429, "lpips": 0.07673 } }, "chunk": { "view": { "psnr": 24.2875, "ssim": 0.81666, "lpips": 0.07277 }, "tl": { "psnr": 49.7579, "ssim": 0.99322, "lpips": 0.00724 }, "tr": { "psnr": 40.5555, "ssim": 0.96978, "lpips": 0.06348 } }, "noisolate": { "view": { "psnr": 23.3697, "ssim": 0.79943, "lpips": 0.07906 }, "tl": { "psnr": 49.964, "ssim": 0.99333, "lpips": 0.0073 }, "tr": { "psnr": 39.2103, "ssim": 0.96396, "lpips": 0.07524 } } } }, { "sample": "sample_023", "episode": "motherboard_0510_episode_006.pt", "t_start": 1566, "first_pred_frame": 5, "metrics": { "delta": { "view": { "psnr": 23.0775, "ssim": 0.78158, "lpips": 0.07877 }, "tl": { "psnr": 49.3075, "ssim": 0.99253, "lpips": 0.00775 }, "tr": { "psnr": 33.8873, "ssim": 0.94809, "lpips": 0.12501 } }, "chunk": { "view": { "psnr": 22.7761, "ssim": 0.78098, "lpips": 0.08714 }, "tl": { "psnr": 49.3288, "ssim": 0.99244, "lpips": 0.00776 }, "tr": { "psnr": 34.3849, "ssim": 0.95275, "lpips": 0.1511 } }, "noisolate": { "view": { "psnr": 22.1697, "ssim": 0.77313, "lpips": 0.09779 }, "tl": { "psnr": 49.7369, "ssim": 0.99269, "lpips": 0.00775 }, "tr": { "psnr": 34.814, "ssim": 0.951, "lpips": 0.12084 } } } } ], "probe": { "rows": [ { "model": "a9dfilm_isolate_fps15", "all": 5.24156139249393, "trans": 6.048608124054235, "rot": 4.623417542733266, "n": 72, "hl": false }, { "model": "dforce_delta_fps15", "all": 5.245501504649816, "trans": 5.496748470011644, "rot": 4.929571771852032, "n": 72, "hl": true }, { "model": "dforce_chunk_fps15", "all": 6.145055975215692, "trans": 6.94608168472537, "rot": 5.794825691069963, "n": 72, "hl": true }, { "model": "wan5b_vmx", "all": 6.820993473810317, "trans": 7.359715070041615, "rot": 6.640094266268522, "n": 72, "hl": false }, { "model": "dforce_noisolate_fps15", "all": 6.984157711832033, "trans": 6.864965269826607, "rot": 7.063216639848914, "n": 72, "hl": true }, { "model": "a9dfilm_fps15", "all": 7.324126243795831, "trans": 7.954014444824645, "rot": 7.045984472272579, "n": 72, "hl": false }, { "model": "aimotion", "all": 8.258736775837962, "trans": 9.085174534106638, "rot": 5.954317444902641, "n": 72, "hl": false }, { "model": "gelz", "all": 10.085044564277691, "trans": 9.914008291742661, "rot": 10.095698106160995, "n": 72, "hl": false }, { "model": "gelzF", "all": 10.405379209914805, "trans": 10.803954730301164, "rot": 10.183212687096717, "n": 72, "hl": false }, { "model": "gelz_jitv", "all": 10.754167474261994, "trans": 10.594921144738326, "rot": 12.152321467208576, "n": 72, "hl": false }, { "model": "gelz_jit", "all": 11.05161020331287, "trans": 13.076652824230079, "rot": 9.474547521704462, "n": 72, "hl": false }, { "model": "a9dpixflow", "all": 11.305266392230571, "trans": 17.589166409564083, "rot": 7.439902752467703, "n": 72, "hl": false }, { "model": "gelnoF", "all": 11.48036148470494, "trans": 12.50174121704579, "rot": 10.720640289067862, "n": 72, "hl": false }, { "model": "a9dfilm", "all": 12.466668534348521, "trans": 13.040619197036108, "rot": 11.897830995778977, "n": 72, "hl": false }, { "model": "gelF", "all": 13.866101828751605, "trans": 13.891061561531655, "rot": 13.866101828751605, "n": 72, "hl": false }, { "model": "a9dfilm_isolate", "all": 14.687242374900606, "trans": 14.71178700072263, "rot": 14.687242374900606, "n": 72, "hl": false }, { "model": "gelz_hist", "all": 16.011046508034084, "trans": 13.780634660653924, "rot": 16.956675106067426, "n": 72, "hl": false } ], "pairs": [ { "a": "dforce_delta_fps15", "b": "dforce_chunk_fps15", "all": { "n": 72, "median": -1.3116269323562968, "wins": 53, "p": 7.555811484431396e-05, "a": 5.245501504649816, "b": 6.145055975215692 }, "trans": { "n": 36, "median": -2.2652683953032384, "wins": 28, "p": 0.0011932429624721408, "a": 5.496748470011644, "b": 6.94608168472537 }, "rot": { "n": 36, "median": -0.8625600155135884, "wins": 25, "p": 0.028816719655878842, "a": 4.929571771852032, "b": 5.794825691069963 }, "runs": { "0": { "n": 12, "median": -2.228386865543139, "wins": 9, "p": 0.14599609375, "a": 5.6835861087535235, "b": 10.339709685000507 }, "1": { "n": 12, "median": -0.45043635166933127, "wins": 8, "p": 0.3876953125, "a": 5.800509729174208, "b": 7.600540206910532 }, "2": { "n": 12, "median": -2.888250582566119, "wins": 10, "p": 0.03857421875, "a": 3.4212059820458522, "b": 5.838904213568022 }, "3": { "n": 12, "median": -1.8519421387485397, "wins": 11, "p": 0.00634765625, "a": 2.3179310684290044, "b": 5.026677519221099 }, "4": { "n": 12, "median": -1.283557732579828, "wins": 9, "p": 0.14599609375, "a": 12.319647867492531, "b": 15.393194855924587 }, "5": { "n": 12, "median": 0.5194583649894791, "wins": 6, "p": 1.0, "a": 4.666715917626309, "b": 3.811945583095117 } }, "title": "delta_action vs action_chunking (the action representation)", "line": "data.action_repr", "a_label": "delta_action + isolate", "b_label": "action_chunking + isolate" }, { "a": "dforce_delta_fps15", "b": "dforce_noisolate_fps15", "all": { "n": 72, "median": -0.572607464120025, "wins": 41, "p": 0.28878428041410487, "a": 5.245501504649816, "b": 6.984157711832033 }, "trans": { "n": 36, "median": -0.8336529561546118, "wins": 22, "p": 0.2429849540349096, "a": 5.496748470011644, "b": 6.864965269826607 }, "rot": { "n": 36, "median": -0.02830317712130287, "wins": 19, "p": 0.8679394004284404, "a": 4.929571771852032, "b": 7.063216639848914 }, "runs": { "0": { "n": 12, "median": -3.036514013419933, "wins": 11, "p": 0.00634765625, "a": 5.6835861087535235, "b": 9.858984594453315 }, "1": { "n": 12, "median": -0.5457524930651716, "wins": 7, "p": 0.7744140625, "a": 5.800509729174208, "b": 7.629233074890825 }, "2": { "n": 12, "median": -1.1711654143403654, "wins": 8, "p": 0.3876953125, "a": 3.4212059820458522, "b": 4.670578280567762 }, "3": { "n": 12, "median": -3.1314645567478934, "wins": 10, "p": 0.03857421875, "a": 2.3179310684290044, "b": 6.133644409478805 }, "4": { "n": 12, "median": 1.8775548781807263, "wins": 2, "p": 0.03857421875, "a": 12.319647867492531, "b": 11.46528591319611 }, "5": { "n": 12, "median": 1.0405645574619888, "wins": 3, "p": 0.14599609375, "a": 4.666715917626309, "b": 3.547993966650809 } }, "title": "isolate_sensors true vs false (do the two gels attend each other?)", "line": "model.isolate_sensors", "a_label": "delta_action + isolate", "b_label": "delta_action, no isolate" } ] }, "lede": "Three runs encode the tactile force as a change of indentation (data.force_term: delta, k_f·(f[t+1]−f[t])) on the 15 fps latent tree, with no pixel-flow branch. They form two one-line comparisons that share a corner: data.action_repr (per-step delta_action vs anchor-relative action_chunking) and model.isolate_sensors (may the two gels attend each other?). Same data, same seed, and the held-out rollouts below draw the same 24 windows in every arm (asserted, not assumed).", "verdict": "The action representation splits by stream, and the cross-sensor attention mask costs tactile accuracy. delta_action beats action_chunking on the camera view (24.30 vs 23.99 dB, 18–20 of 24 windows) and on commanded actions (5.5 vs 6.9 px median translation error), while action_chunking wins the tactile images (right gel 42.23 vs 41.47 dB, 5 of 24). Turning the isolation mask OFF — letting the two gels attend each other again — improves both gels (right gel 42.30 vs 41.47 dB; SSIM wins 20/24 and 19/24 for the unmasked arm, p≤0.007) at no measurable cost to the view (15/24, p=0.31). This is the first time the isolate ablation has moved anything — it tied on val loss at 6 fps and 15 fps with the absolute force term, and it ties again here (0.0901 / 0.0902 / 0.0901). On the probes its median favours the masked arm (5.2 vs 7.0 px) but the paired count is only 41 of 72 (p=0.29) and flips sign on two of the six start frames, so that gap is not supported.", "diff": "data:\n force_term: delta # both arms\n action_repr: delta_action # job 514906\n action_repr: action_chunking # job 514907\n\ndelta_action row k = pose step idx[k] -> idx[k+1], force f[k+1] - f[k]\naction_chunking row k = pose idx[0] -> idx[k+1], force f[k+1] - f[0]\n (idx[0] = the window's first frame, the anchor)", "diff_note": "The two are the same converter on a different pose pair; a window's delta rows telescope to the chunk's last row. They also compress differently into the 3 latent transitions the model conditions on: delta_action average-pools the 5 frame rows of a block, action_chunking takes the block's LAST row (averaging anchor-relative rows would be meaningless).", "loss_note": "Both ran 100 epochs on 2xH100 and overfit after ~ep35, so these are the best-* checkpoints, not last.ckpt: final-epoch val is 0.0950 / 0.0949. The run's split file warns that val aliases test, so treat every number here as a development metric.", "metric_note": "Predicted frames only. The rollout is given 2 context latent frames (decoded frames 0–4) and generates 2 more (frames 5–12); averaging over all 13 frames, as the older decode script does, adds ~2 dB of free PSNR from frames that are identical by construction. Green = better of the two arms.", "sign_note": "Paired over the same 24 held-out windows. Windows inside one episode overlap in time, so they are not fully independent and the p-values are optimistic; the win counts are the more honest summary. 24 windows can at best reach p=1.2e-7.", "probe_note": "React probes_v1: 72 commanded axis-aligned motions from 6 real start frames, scored geometrically by tracking the gel centre through the decoded video with AllTracker (median final error, crop px; lower is better). All models on one table, but the 15 fps rows measure a 0.53 s horizon and the 6 fps rows 1.33 s — compare within a rate only. Do-nothing scores 38.7 px on translations.", "pair_note": "The 72 probes are 6 start frames x 12 directions, and probes sharing a start frame share context and episode, so the 'all 72' p-value treats correlated samples as independent. Read the per-start-frame rows: delta is ahead on 5 of 6.", "strip_note": "Row 1 is the model's prediction, row 2 the ground truth, one column per decoded frame. Frames 0–4 come from the 2 given context latents (they are GT in both rows); the red box marks frame 5, the first generated one.", "caveats": [ "The isolate comparison is a mask-only change: no parameter count or weight shape changes, so the two arms differ solely in which attention edges exist. The same ablation has come out a tie on val loss twice before (fps6 and fps15, absolute force); this is the third setting it is asked in.", "The delta arms' force channel is dead on these probes. With the commanded future force held constant, every per-step row has Δf = 0, so act_wdforce_hold IS the plain rigid action. The chunk arm still sees f_commanded − f_anchor. The two arms are not on equal footing on the force channel — that is a property of the encodings, not of the eval.", "The absolute-force reference rows are not paired with the dforce cells: they were rolled out on 10 windows drawn independently, and random.sample(N, 24) does not extend random.sample(N, 10).", "Only the two named pairs are controlled. action_chunking+isolate against delta_action-without-isolate differs in two lines at once; the summary table shows all three arms side by side, but read a difference only along one axis.", "Probe grid offset. For start frames run1 and run4 the probe's commanded pose p[0] sits one native frame off the window grid, up to 10.3 mm from the real pose at t_ctx. The spliced per-step stream silently drops that jump; the anchor-relative rows carry it. Pre-existing, and it hits the absolute-force runs identically.", "Bug fixed to make this eval valid: scripts/inference_rollout_mot.py called _compress_action(act, W_lat) without action_repr, so every action_chunking run was rolled out with average-pooled anchor-relative rows instead of block-last ones. Any earlier short-horizon eval of a chunk run is suspect.", "Single seed (0) for both the rollouts and the probes." ], "repro": "# held-out rollout + PSNR/SSIM/LPIPS + PRED-over-GT strips; ARMS_SPEC selects the\n# arms and the cells are idempotent, so an added arm reuses the existing ones\nsbatch vm_diffusion/run_eval_dforce_ab.sh\n\n# commanded rows for the delta encoding (act_wdforce_* / act_chunk_wdforce_*)\nPROBE_FPS=15 python -m vm_diffusion.probe_eval.add_dforce_actions\n\n# React probe eval (MODELS_SPEC picks the arms; default = the first two)\nsbatch vm_diffusion/probe_eval/run_probe_eval_dforce.sh\n\n# this page\npython -m vm_diffusion.scripts.build_hf_space_dforce --push\n" }