{ "schema_version": "paper_public_manifest_v1", "paper_id": "pegasus-editable-experience_2026_08", "slug": "pegasus-editable-experience", "title": "Cross-embodiment transfer should be done at the experience level", "authors": [], "source": { "arxiv_id": "2607.26903", "pdf_url": "https://arxiv.org/pdf/2607.26903", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2607.26903" }, "site": { "post_url": "/posts/pegasus-editable-experience", "canonical_url": "https://haiguangboy.com/posts/pegasus-editable-experience", "cover_image": "https://static.haiguangboy.com/papers/pegasus-editable-experience/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "robotics", "tasks": [ "embodied_ai", "robotics", "state_prediction", "Embodied intelligence", "Cross-embodiment transfer", "Robot data" ], "related_topics": [ { "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": "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": "decompose_and_reorganize_planning_with_primitives_and_visuomotor_policies_learne_2026_08", "title": "Switching logic shouldn't be learned by the policy", "url": "https://haiguangboy.com/posts/dr-lfd", "relation": "same_track", "summary": "The graph translation is decomposed into a cascade of three operators: affordance graph → constraint graph → planning graph, validating the positioning of foundation models: VLM semantics are only useful when converted into verifiable action structures", "strength": "strong" }, { "paper_id": "archon_blog_whole_body_intelligence_cn_20260712_2026_07", "title": "Li Hongyang's whole-body intelligence surpasses GR's whole-body control", "url": "https://haiguangboy.com/posts/archon-whole-body-intelligence", "relation": "same_track", "summary": "Key design: choosing affordance rather than object identity or pose as the cross-embodiment invariant, organized in three layers, validating the evaluation criterion: don't look at single demonstrations, look at the speed of capability improvement on new scenes/new hardware/new tasks", "strength": "strong" }, { "paper_id": "wx_elsewhere别处发生_20260722_2026_07", "title": "Transcript of Liang Wenfeng's four-hour investor meeting", "url": "https://haiguangboy.com/posts/liangwenfeng-world-model", "relation": "same_track", "summary": "Five-stage pipeline: task graph → hierarchical affordance latent variables → cross-embodiment graph translation → physical constraint verification → experience rendering, validating ★ The disagreement may stem from different referents: the \"3D/video generation/world model\" that Liang Wenfeng lists in parallel all fall into the \"renderer\" category, while robots need a \"simulator\"", "strength": "strong" }, { "paper_id": "leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07", "title": "Predictive features cannot be fed directly to diffusion models", "url": "https://haiguangboy.com/posts/leapbot-wa", "relation": "same_track", "summary": "Five-stage pipeline: task graph → hierarchical affordance latent variables → cross-embodiment graph translation → physical constraint verification → experience rendering, validating key design: the world model only looks at intent, not actions—deliberately depriving it of low-level motor commands to prevent collapse into a control shortcut", "strength": "strong" } ] }, "ruling": { "importance_score": 3.0, "one_sentence": "Cross-embodiment transfer should be done at the experience level" }, "asset_base_url": "https://static.haiguangboy.com/papers/pegasus-editable-experience", "assets": [ { "type": "cover_image", "object_key": "papers/pegasus-editable-experience/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pegasus-editable-experience/cover.webp", "role": "post_cover", "size_bytes": 92538 }, { "type": "public_brief", "object_key": "papers/pegasus-editable-experience/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pegasus-editable-experience/public_brief.md", "role": "public_brief", "size_bytes": 5492 }, { "type": "public_manifest", "object_key": "papers/pegasus-editable-experience/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pegasus-editable-experience/public_manifest.json", "role": "public_manifest", "size_bytes": 5576 } ], "published_at": "2026-08-04T16:45:50+08:00", "created_at": "2026-08-04T16:45:50+08:00", "updated_at": "2026-09-02T10:55:06+08:00", "analyst_take": { "type": "author_opinion", "text": "Setting the numbers aside and looking at the methodology, there is one genuine disagreement worth noting. Another line of work argues: rather than making the policy generalize to unseen situations, it is better to first use primitives to push the world back into the policy's distribution—what changes is not the model's generalization ability, but the state it is in when invoked. This work takes the opposite view, arguing for using affordance combinations to let the model itself infer how to manipulate unseen objects.\n\nBoth paths answer the same question, but the answers point to opposite ends. One keeps uncertainty outside the policy; the other requires the policy to possess compositional reasoning. The former is more controllable in engineering, while the latter has a higher ceiling but is harder to verify—and this work precisely lacks real-robot execution to support the latter." } }