{ "schema_version": "paper_public_manifest_v1", "paper_id": "splatting-physical-scenes_2026_08", "slug": "splatting-physical-scenes", "title": "Scale is not given by the model, but by the robot", "authors": [], "source": { "arxiv_id": "2506.04120", "pdf_url": "https://arxiv.org/pdf/2506.04120", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2506.04120" }, "site": { "post_url": "/posts/splatting-physical-scenes", "canonical_url": "https://haiguangboy.com/posts/splatting-physical-scenes", "cover_image": "https://static.haiguangboy.com/papers/splatting-physical-scenes/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "robotics", "tasks": [ "embodied_ai", "robotics", "state_prediction", "Embodied Intelligence", "Robot Learning", "3dgs", "real2sim" ], "related_topics": [ { "paper_id": "deng_video2robo_3dgs_based_synthetic_data_from_one_video_enables_scalable_robot_2026_08", "title": "Closed is the kinematic loop, not the physical loop", "url": "https://haiguangboy.com/posts/video2robo", "relation": "contrast", "summary": "Problem setup: directly producing simulation assets from imperfect data of low-cost robots, rather than first requiring a clean collection environment contradicts This is not physical simulation: the entire pipeline is kinematic, without contact forces, friction, or dynamic models", "strength": "strong" }, { "paper_id": "leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07", "title": "Predictive features cannot be directly fed into diffusion models", "url": "https://haiguangboy.com/posts/leapbot-wa", "relation": "contrast", "summary": "Two representation hard limits: MJX only supports rigid bodies, and the 3DGS rendering model cannot be re-lit contradicts Problem diagnosis: mainstream WAM inherits from video generation, treating world modeling as a pixel-level rendering problem", "strength": "strong" }, { "paper_id": "d4rt_efficiently_reconstructing_dynamic_scenes_one_2026_06", "title": "D4RT_Efficiently_Reconstructing_Dynamic_Scenes_One", "url": "https://haiguangboy.com/posts/d4rt_efficiently_reconstructing_dynamic_scenes_one", "relation": "same_track", "summary": "Key difference compared to TRELLIS: this method outputs metric scale and 6D pose, while generative models do not validates D4RT output is relative/up-to-scale, not metric—this is the critical issue when used as an embodied 3D data source", "strength": "strong" }, { "paper_id": "transformer-transformer_blog_transformer_transformer_a_unified_20260807_2026_08", "title": "Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design", "url": "https://haiguangboy.com/posts/transformer-transformer-codesign", "relation": "same_track", "summary": "Two representation hard limits: MJX only supports rigid bodies, and the 3DGS rendering model cannot be re-lit validates The design space lacks complex meshes, scenes, and contact targets, leaving key gaps for assembly and dexterous contact", "strength": "strong" }, { "paper_id": "dexora_open_source_vla_for_high_dof_bimanual_dexterity_2026_08", "title": "The noise in teleoperation data must enter the loss function", "url": "https://haiguangboy.com/posts/dexora", "relation": "same_track", "summary": "Route claim: the imprecision of low-cost robots is not an obstacle to bypass, but an object to be corrected in place using visual feedback validates Core engineering claim: the noise in teleoperation data must be explicitly modeled, not diluted by scale", "strength": "strong" }, { "paper_id": "causally_debiased_latent_action_model_for_embodied_action_conditioned_world_mode_2026_07", "title": "The world model is disobedient because latent actions are contaminated", "url": "https://haiguangboy.com/posts/cd-lam", "relation": "contrast", "summary": "Engineering finding: COLMAP fails to produce consistent solutions on this data, and SAM2 cannot segment the robot body either contradicts Method: three debiasing objectives—ego-centric reconstruction, action-centric contrast, and latent space calibration", "strength": "medium" } ] }, "ruling": { "importance_score": 3.0, "one_sentence": "Let the robot use itself as a ruler; reconstruction error is 3.31mm, while generative models are nearly 5 times worse" }, "asset_base_url": "https://static.haiguangboy.com/papers/splatting-physical-scenes", "assets": [ { "type": "public_brief", "object_key": "papers/splatting-physical-scenes/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/splatting-physical-scenes/public_brief.md", "role": "public_brief", "size_bytes": 5582 }, { "type": "public_manifest", "object_key": "papers/splatting-physical-scenes/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/splatting-physical-scenes/public_manifest.json", "role": "public_manifest", "size_bytes": 5372 } ], "published_at": "2026-08-10T16:07:47+08:00", "created_at": "2026-08-10T16:07:47+08:00", "updated_at": "2026-09-02T10:55:06+08:00", "analyst_take": { "type": "author_opinion", "text": "The previous Video2Robo bet was to bypass physics using kinematics plus photorealistic rendering; this paper does the opposite. Interestingly, the same judgment pulls two papers in opposite directions: LeapBot-WA argues that the utility of world modeling lies in abstract physical prediction rather than photorealistic rendering—Video2Robo goes against it, while this paper confirms it.\n\nThe scale thread ties it together. D4RT's output is up-to-scale, not metric; Video2Robo relies on VGGT's monocular depth to compensate; this paper integrates MJX's kinematics into the optimization loop, where the robot's own mechanical dimensions serve as the ruler—so it dares to claim metrically accurate, while TRELLIS does not even provide scale or pose.\n\nBut both papers share the same ceiling: they only support rigid bodies.\n\nAnchor point for review in six months: a calibration error of 0.01 radians can be corrected to 3.79mm, while 0.03 radians only reaches 18.5mm, and the authors do not discuss this degradation. Whether contact-intensive tasks can tolerate residual errors of one to two centimeters is the watershed for whether this route can truly serve as a physical foundation." } }