{ "schema_version": "paper_public_manifest_v1", "paper_id": "learning_native_continuation_for_action_chunking_flow_policies_2026_09", "slug": "learning_native_continuation_for_action_chunking_flow_policies", "title": "VLA换动作块不再卡顿", "authors": [], "source": { "arxiv_id": "2602.12978", "pdf_url": "https://arxiv.org/pdf/2602.12978", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/pdf/2602.12978" }, "site": { "post_url": "/posts/learning_native_continuation_for_action_chunking_flow_policies", "canonical_url": "https://haiguangboy.com/posts/learning_native_continuation_for_action_chunking_flow_policies", "cover_image": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "vla", "tasks": [ "embodied_ai", "vla", "action_generation", "robotics", "具身智能", "动作分块" ], "related_topics": [ { "paper_id": "an_open_foundation_model_towards_2026_07", "title": "An_Open_Foundation_Model_Towards", "url": "https://haiguangboy.com/posts/an_open_foundation_model_towards", "relation": "contrast", "summary": "核心路线判断:块间衔接应该是策略自身学到的去噪动力学里的原生属性,不该是套在标准速度场外面、靠推理时或训练时硬约束打补丁的外部机制——这是Legato区别于RTC和训练时RTC的根本立场 contradicts 训练时RTC：遮掩前d个动作token使模型学会平滑续接", "strength": "strong" }, { "paper_id": "pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08", "title": "PointWorld:点流统一状态动作", "url": "https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation", "relation": "same_track", "summary": "PointWorld:点流统一状态动作", "strength": "strong" }, { "paper_id": "beyond_action_residuals_real_world_robot_policy_steering_via_bottleneck_latent_r_2026_08", "title": "RL 该在哪一层介入", "url": "https://haiguangboy.com/posts/zprl", "relation": "same_track", "summary": "跟同期出现的'训练时RTC'(也在训练阶段引入衔接约束,但只是给重叠动作段加一个硬性前缀约束、不改变底层flow动力学本身)对比,Legato在任务分数、完成时间、平滑度上全面胜出——说明真正重要的是重塑策略本身的去噪动力学,不是单纯加一个训练时硬约束 validates 问题重述:RL 后训练的关键不只是「改多少」,更是「在哪一层介入」", "strength": "strong" }, { "paper_id": "wx_中国财富_20260719_2026_07", "title": "千寻智能韩峰涛：具身智能进入“拼速度”阶段", "url": "https://haiguangboy.com/posts/qianxun-general-brain-data-loop", "relation": "same_track", "summary": "论文作者在千寻智能(Spirit AI)实习期间完成此工作(团队负责人为高阳),被机器人顶会RSS 2026接收;Legato得名于音乐术语「连音」——音符间平滑过渡、毫无间断 validates ★团队组合：产业老兵+Abbeel门生，高阳师承直连Physical Intelligence/Covariant创始人", "strength": "strong" }, { "paper_id": "flex_pi_a_multi_stream_world_action_model_with_compute_flexibility_2026_08", "title": "把RGB、3D几何、物体语义揉进同一个表征——这才是它比π0.5、Fast-WAM更准的根本原因", "url": "https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility", "relation": "same_track", "summary": "FLEX-π:更全面的联合表征", "strength": "strong" }, { "paper_id": "omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08", "title": "未来预测越准，机器人反而越迟滞", "url": "https://haiguangboy.com/posts/omega-0", "relation": "same_track", "summary": "动作分块(chunking)是VLA部署的标配,能摊薄推理成本、支持高频控制,但推理延迟加上flow策略内在的多模态性,导致相邻两个动作块衔接处经常不平滑,表现为犹豫和突兀转向,拖长任务完成时间 validates 部署:单次前向约 0.14 秒(超过 7Hz),预测 25 步动作块只执行前 8 步,RTC 式热启动加重叠混合平滑衔接", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "\"训练时把部署期会变化的麻烦参数随机化、让策略自己适应\"这个手法，和已解读过的PointWorld同源——PointWorld随机化相机数量、这篇随机化推理延迟，结果都是模型对该参数的各种取值反而最鲁棒。调度消融里步幅和坡降长度的权衡，也和PAVE的预测锚点数量非单调发现是同一个模式：给系统更多约束或参数不是无条件的好事，关键在找到那个恰好的权衡点。" }, "ruling": { "importance_score": 3.0, "one_sentence": "VLA换动作块不再卡顿" }, "asset_base_url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies", "assets": [ { "type": "pdf_screenshot", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/page_01.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/page_01.webp", "role": "paper_first_page", "size_bytes": 237606 }, { "type": "pdf_screenshot", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/key_figure.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/key_figure.webp", "role": "method_figure", "size_bytes": 230864 }, { "type": "cover_image", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/cover.webp", "role": "post_cover", "size_bytes": 83484 }, { "type": "public_brief", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.md", "role": "public_brief", "size_bytes": 5200 }, { "type": "public_brief", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.en.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.en.md", "role": "public_brief_en", "size_bytes": 6844 }, { "type": "public_manifest", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.en.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.en.json", "role": "public_manifest_en", "size_bytes": 8569 }, { "type": "public_manifest", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.json", "role": "public_manifest", "size_bytes": 8896 } ], "published_at": "2026-09-12T22:48:17+08:00", "created_at": "2026-09-12T22:48:17+08:00", "updated_at": "2026-09-12T22:48:29+08:00" }