{
  "schema_version": "paper_public_manifest_v1",
  "paper_id": "tacwam_2026_08",
  "slug": "tacwam",
  "title": "Tactile prediction should only be used for training, not shown to actions",
  "authors": [],
  "source": {
    "arxiv_id": "2607.28391",
    "pdf_url": "https://arxiv.org/pdf/2607.28391",
    "project_url": "",
    "github_url": "",
    "huggingface_url": "",
    "original_source": "https://arxiv.org/abs/2607.28391"
  },
  "site": {
    "post_url": "/posts/tacwam",
    "canonical_url": "https://haiguangboy.com/posts/tacwam",
    "cover_image": "https://static.haiguangboy.com/papers/tacwam/cover.webp"
  },
  "taxonomy": {
    "domain": "embodied_ai",
    "track": "world_model",
    "tasks": [
      "embodied_ai",
      "world_model",
      "vla",
      "action_generation",
      "robotics",
      "latent_state",
      "state_prediction",
      "Embodied intelligence",
      "Tactile perception",
      "World models"
    ],
    "related_topics": [
      {
        "paper_id": "n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens_2026_07",
        "title": "Tactile as prediction target rather than observation input",
        "url": "https://haiguangboy.com/posts/n_0_vtla_scaling_vision_tactile_language_action_model_with_latent_tactile_tokens",
        "relation": "contrast",
        "summary": "Tactile as prediction target rather than observation input",
        "strength": "strong"
      },
      {
        "paper_id": "latent_action_pretraining_through_world_modeling_2026_07",
        "title": "LAWM: Why action labels become a burden",
        "url": "https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling",
        "relation": "same_track",
        "summary": "LAWM: Why action labels become a burden",
        "strength": "strong"
      },
      {
        "paper_id": "t_rex_tactile_reactive_dexterous_manipulation_2026_07",
        "title": "T-Rex: Why tactile needs separate modeling",
        "url": "https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation",
        "relation": "same_track",
        "summary": "T-Rex: Why tactile needs separate modeling",
        "strength": "strong"
      },
      {
        "paper_id": "orca_2026_07",
        "title": "π0.5 trembles in place when failing to grab a spoon, but Orca's physical intuition learned from watching videos goes further",
        "url": "https://haiguangboy.com/posts/orca",
        "relation": "same_track",
        "summary": "The key to a world model is readable states",
        "strength": "strong"
      },
      {
        "paper_id": "robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode_2026_07",
        "title": "It's not 'whether to have a world model', it's 'whether what it outputs has structure'",
        "url": "https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode",
        "relation": "same_track",
        "summary": "It's not 'whether to have a world model', it's 'whether what it outputs has structure'",
        "strength": "strong"
      },
      {
        "paper_id": "pose_aware_modeling_to_mitigate_pose_related_artifacts_in_tactile_gloves_2026_07",
        "title": "Glove tactile readings are mixed with hand pose",
        "url": "https://haiguangboy.com/posts/tactile-glove-pose-artifacts",
        "relation": "same_track",
        "summary": "Very small scale: single robotic arm platform, four tasks, 300 demos per task, 20 trials per method; weak statistical foundation: only 5 people per glove, p-values discretized at 0.03125; no zero-shot cross-glove transfer",
        "strength": "strong"
      }
    ]
  },
  "ruling": {
    "importance_score": 3.0,
    "one_sentence": "Tactile prediction should only be used for training, not shown to actions"
  },
  "asset_base_url": "https://static.haiguangboy.com/papers/tacwam",
  "assets": [
    {
      "type": "cover_image",
      "object_key": "papers/tacwam/cover.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/tacwam/cover.webp",
      "role": "post_cover",
      "size_bytes": 91460
    },
    {
      "type": "public_brief",
      "object_key": "papers/tacwam/public_brief.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/tacwam/public_brief.md",
      "role": "public_brief",
      "size_bytes": 5277
    },
    {
      "type": "public_manifest",
      "object_key": "papers/tacwam/public_manifest.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/tacwam/public_manifest.json",
      "role": "public_manifest",
      "size_bytes": 4928
    }
  ],
  "published_at": "2026-08-03T18:18:19+08:00",
  "created_at": "2026-08-03T18:18:19+08:00",
  "updated_at": "2026-09-02T10:55:06+08:00",
  "analyst_take": {
    "type": "author_opinion",
    "text": "That ablation curve deserves to be highlighted on its own. The last two ablation levels did not reduce any available information, only removed structural constraints on the information flow, yet the loss was greater than the information loss in the first level. The judgment that 'auxiliary signals cannot be plugged into the backbone without constraints' has previously only been supported by positive evidence that 'adding constraints works better'; this is the first time seeing a descending curve as constraints are progressively relaxed.\n\nThere is also an interesting divergence within the same school. Another tactile work also advocates treating tactile as a prediction target rather than observation input, but it feeds the predicted latent variables to an action expert to condition actions; this paper explicitly does not let actions read tactile futures. Both start from the same point but give opposite answers to 'whether predicted outputs should be shown to actions'. And this paper's ablation happens to test exactly that, dropping 17.5 points."
  }
}