analysis/ablation_meta.json

{
  "01_vo_right_single_view_BEST_vis 0.0093": {
    "category": "A. Single-view vision-only baselines",
    "short": "vo_right (single view, no tactile)",
    "blurb": "Visual-only diffusion world model trained on the right RGB camera. No tactile streams, no shift, no delta-ref. Best single-view visual baseline.",
    "setup": {
      "fusion": "channel",
      "views": "1 (right)",
      "tactile": "\u2014",
      "shift16": "\u2717",
      "delta-ref": "\u2014",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0093"
    }
  },
  "02_p3_mv2t_3v2t_no_delta_no_shift val 0.0139": {
    "category": "B. Multi-view + tactile (no fixes \u2014 old baselines)",
    "short": "p3_mv2t (3V+2T, NO fixes)",
    "blurb": "MM-DiT-MV with 3 visual streams + 2 tactile streams. No shift, no delta-ref. tactile val=0.0040 looks fine but it's dominated by no-contact frames (trivial baseline).",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (abs)",
      "shift16": "\u2717",
      "delta-ref": "\u2717",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0139"
    }
  },
  "03_p4_mv2t_extr_NO_GATE_NO_DELTA val 0.335 BLOWUP": {
    "category": "B. Multi-view + tactile (no fixes \u2014 old baselines)",
    "short": "p4_mv2t_extr (+ cam-pose, NO gate, NO delta-ref)",
    "blurb": "Adding camera-pose extrinsics directly to the visual AdaLN path breaks tactile: gradients flow through joint attention and poison tactile tokens. Tactile val explodes from 0.004 \u2192 0.32.",
    "badge": "blowup",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (abs)",
      "shift16": "\u2717",
      "delta-ref": "\u2717",
      "cam-pose": "\u2713",
      "gate": "\u2717",
      "val_loss": "0.335 \ud83d\udca5"
    }
  },
  "04_p4_gate_GATE_NO_DELTA val 0.275 tactile_still_bad": {
    "category": "C. Architectural workarounds (gate)",
    "short": "p4_gate (+ scalar gate, no delta-ref)",
    "blurb": "Added a zero-init scalar gate on cam-pose injection (SVD-style soft onset). Saved visual val_loss to 0.0136 (matches single-view best) but tactile is still 0.26 \u2014 gate slows cam-pose entry but cannot prevent cross-modal interference once it opens.",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (abs)",
      "shift16": "\u2717",
      "delta-ref": "\u2717",
      "cam-pose": "\u2713",
      "gate": "\u2713",
      "val_loss": "0.275 (vis 0.0136)"
    }
  },
  "05_vtwm_s1_1v2t_SHIFT_no_delta val 0.0196": {
    "category": "D. Single-view + tactile (shift / delta-ref ablation)",
    "short": "vtwm_s1 (1V2T + shift16, NO delta-ref)",
    "blurb": "1V+2T with tactile-shift=16 frames (compensates ~500 ms sensor latency). Single view, no delta-ref. Shows that shift alone already prevents tactile blowup.",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "1",
      "tactile": "L+R (abs)",
      "shift16": "\u2713",
      "delta-ref": "\u2717",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0196"
    }
  },
  "06_vtwm_s2_1v2t_SHIFT+DELTA val 0.0189": {
    "category": "D. Single-view + tactile (shift / delta-ref ablation)",
    "short": "vtwm_s2 (1V2T + shift16 + delta-ref)",
    "blurb": "Same as s1 but predicts tactile as delta-from-reference instead of absolute latent. Delta-ref gives a small but consistent win over s1 (0.0196 \u2192 0.0189).",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "1",
      "tactile": "L+R (delta)",
      "shift16": "\u2713",
      "delta-ref": "\u2713",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0189"
    }
  },
  "07_mv2_3v2t_CAMPOSE+SHIFT+DELTA val 0.0136": {
    "category": "E. SOTA (multi-view + full recipe)",
    "short": "mv2_3v2t_campose_delta_ref (full fix)",
    "blurb": "3V+2T with cam-pose, shift16 AND delta-ref. Delta-ref makes tactile invariant to absolute-latent perturbation, so cam-pose injection no longer hurts tactile. val 0.0136 vs p4_gate 0.275 = 20\u00d7 better.",
    "badge": "\ud83e\udd47 SOTA",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (delta)",
      "shift16": "\u2713",
      "delta-ref": "\u2713",
      "cam-pose": "\u2713",
      "gate": "\u2713",
      "val_loss": "0.0136"
    }
  },
  "08_mv1_3v2t_SHIFT+DELTA NO_CAMPOSE val 0.0135 SOTA": {
    "category": "E. SOTA (multi-view + full recipe)",
    "short": "mv1_3v2t_delta_ref (no cam-pose)",
    "blurb": "Same as mv2 but without cam-pose. Slightly better (0.0135 vs 0.0136). Difference is within noise \u2192 cam-pose with delta-ref is benign (zero penalty, zero added benefit on the test). Suggests cam-pose only matters when there's a downstream task that needs view-aware geometry.",
    "badge": "\ud83e\udd47 SOTA",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (delta)",
      "shift16": "\u2713",
      "delta-ref": "\u2713",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0135"
    }
  },
  "06b_vtwm_p0_1v2t_SHIFT+DELTA_EPISODE_P01 val 0.0202": {
    "category": "D. Single-view + tactile (shift / delta-ref ablation)",
    "short": "vtwm_p0 (1V2T + shift + delta + episode_p01 ref)",
    "blurb": "Same as s2 but uses the quietest episode-wide frame (p01 percentile) as the delta-ref baseline, instead of window's first frame. Conceptually cleaner (delta = real contact magnitude). Latent val 0.0202 (vs s2 0.0189), but VISUAL LPIPS slightly better (0.103 vs 0.104), VISUAL PSNR clearly better (23.5 vs 23.0). Tactile LPIPS slightly worse but tactile SSIM slightly better \u2014 semantic vs texture trade-off.",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "1",
      "tactile": "L+R (delta-p01)",
      "shift16": "\u2713",
      "delta-ref": "\u2713",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0202"
    }
  },
  "08c_mv1_p01_3v2t_SHIFT+DELTA_EPISODE_P01 val 0.0157": {
    "category": "E. SOTA (multi-view + full recipe)",
    "short": "mv1_p01 (3V2T + shift + delta + episode_p01 ref)",
    "blurb": "Same as mv1 but tactile delta is computed against the episode-wide quietest frame (p01 percentile) instead of window's first frame. Conceptually cleaner reference. Latent val 0.0157 (vs mv1 0.0132), view LPIPS 0.094 (vs 0.091), tactile_L LPIPS 0.068 (vs 0.026 - much worse) \u2014 episode_p01 reference is consistently worse than window-first in our 3V2T setup. Window-first wins both at 1V2T (s2 vs P0) and 3V2T (mv1 vs mv1_p01).",
    "setup": {
      "fusion": "mmdit_mv",
      "views": "3",
      "tactile": "L+R (delta-p01)",
      "shift16": "\u2713",
      "delta-ref": "\u2713",
      "cam-pose": "\u2717",
      "gate": "\u2014",
      "val_loss": "0.0157"
    }
  }
}