{ "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 action chunk switching no longer stutters", "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", "Embodied intelligence", "Action chunking" ], "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": "Core route judgment: inter-chunk transitions should be an intrinsic property of the denoising dynamics learned by the policy itself, not an external mechanism patched on top of the standard velocity field through hard constraints at inference or training time—this is the fundamental stance that distinguishes Legato from RTC and training-time RTC contradicts training-time RTC: masking the first d action tokens so the model learns smooth continuation", "strength": "strong" }, { "paper_id": "pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08", "title": "PointWorld: point flow unifies state and action", "url": "https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation", "relation": "same_track", "summary": "PointWorld: point flow unifies state and action", "strength": "strong" }, { "paper_id": "beyond_action_residuals_real_world_robot_policy_steering_via_bottleneck_latent_r_2026_08", "title": "At which layer should RL intervene", "url": "https://haiguangboy.com/posts/zprl", "relation": "same_track", "summary": "Compared with the contemporaneous 'training-time RTC' (which also introduces transition constraints during training, but merely adds a hard prefix constraint to overlapping action segments without changing the underlying flow dynamics themselves), Legato wins across the board on task score, completion time, and smoothness—showing that what truly matters is reshaping the policy's own denoising dynamics, not simply adding a training-time hard constraint validates problem restatement: the key to RL post-training is not just 'how much to change,' but 'at which layer to intervene'", "strength": "strong" }, { "paper_id": "wx_中国财富_20260719_2026_07", "title": "Spirit AI's Han Fengtao: embodied intelligence enters the 'racing for speed' stage", "url": "https://haiguangboy.com/posts/qianxun-general-brain-data-loop", "relation": "same_track", "summary": "The paper authors completed this work during an internship at Spirit AI (team lead: Gao Yang), and it was accepted to the top robotics conference RSS 2026; Legato takes its name from the musical term 'legato'—smooth transitions between notes with no interruption validates ★Team combination: industry veteran + Abbeel protege, Gao Yang's lineage directly connected to the founders of Physical Intelligence/Covariant", "strength": "strong" }, { "paper_id": "flex_pi_a_multi_stream_world_action_model_with_compute_flexibility_2026_08", "title": "Blending RGB, 3D geometry, and object semantics into a single representation—this is the fundamental reason it is more accurate than π0.5 and Fast-WAM", "url": "https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility", "relation": "same_track", "summary": "FLEX-π: a more comprehensive joint representation", "strength": "strong" }, { "paper_id": "omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08", "title": "The more accurate the future prediction, the more sluggish the robot becomes", "url": "https://haiguangboy.com/posts/omega-0", "relation": "same_track", "summary": "Action chunking is standard in VLA deployment, amortizing inference cost and supporting high-frequency control, but inference latency plus the intrinsic multimodality of flow policies often makes the transition between adjacent action chunks unsmooth, manifesting as hesitation and abrupt turns that prolong task completion time validates deployment: a single forward pass takes about 0.14 seconds (over 7Hz), a predicted 25-step action chunk executes only the first 8 steps, and RTC-style warm starting with overlapping blending smooths the transition", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "\"Randomize the troublesome parameters that change during deployment at training time and let the policy adapt on its own\"—this technique shares the same origin as the previously analyzed PointWorld: PointWorld randomizes the number of cameras, this paper randomizes inference latency, and the result in both cases is that the model becomes most robust to various values of that parameter. The trade-off between stride and ramp length in the schedule ablation follows the same pattern as PAVE's non-monotonic finding on the number of prediction anchors: giving a system more constraints or parameters is not unconditionally good; the key is finding that just-right trade-off point." }, "ruling": { "importance_score": 3.0, "one_sentence": "VLA action chunk switching no longer stutters" }, "asset_base_url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies", "assets": [ { "type": "cover_image", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/cover.webp", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/cover.webp", "content_type": "image/webp", "upload_status": "pending", "role": "post_cover" }, { "type": "pdf_screenshot", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/page_01.webp", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/page_01.webp", "content_type": "image/webp", "upload_status": "pending", "role": "paper_first_page" }, { "type": "pdf_screenshot", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/key_figure.webp", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/key_figure.webp", "content_type": "image/webp", "upload_status": "pending", "role": "method_figure" }, { "type": "public_manifest", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.json", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "pending", "role": "public_manifest" }, { "type": "public_brief", "object_key": "papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.md", "url": "https://static.haiguangboy.com/papers/learning_native_continuation_for_action_chunking_flow_policies/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "pending", "role": "public_brief" } ], "published_at": "2026-09-12T22:48:17+08:00", "created_at": "2026-09-12T22:48:17+08:00", "updated_at": "2026-09-12T22:48:17+08:00" }