{ "schema_version": "paper_public_manifest_v1", "paper_id": "dexora_2026_08", "slug": "dexora", "title": "Open-Source VLA with 36 DOF Dual Arms and Dual Dexterous Hands: 66.7% on Dexterous Tasks, 15 Points Higher than GR00T N1", "authors": [], "source": { "arxiv_id": "2605.18722", "pdf_url": "https://arxiv.org/pdf/2605.18722", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2605.18722" }, "site": { "post_url": "/posts/dexora", "canonical_url": "https://haiguangboy.com/posts/dexora", "cover_image": "https://static.haiguangboy.com/papers/dexora/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "vla", "tasks": [ "embodied_ai", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "Embodied Intelligence", "Dexterous Manipulation" ], "related_topics": [ { "paper_id": "wx_晚点latepost_20260718_2026_07", "title": "Sudu Technology Debuts at WAIC: Reality Is the Endpoint for Robots, Not the Boundary for Training", "url": "https://haiguangboy.com/posts/sudu-third-position", "relation": "contrast", "summary": "The gap between simulation and dexterity: simulation can guide basic skills, but dexterity must rely on real-robot data contradicts the precise claim of a simulation-first training route: pretraining uses almost no real-robot data, but after deployment there is failure retry plus a small amount of real-robot RL", "strength": "strong" }, { "paper_id": "deepmind_blog_gemini_robotics_2_brings_20260730_2026_07", "title": "The Overlooked Non-Consensus Is Whole-Body Unified Control", "url": "https://haiguangboy.com/posts/gemini-robotics-2", "relation": "contrast", "summary": "Positioning: existing VLAs are either dual-arm low-DOF grippers or single-arm dexterous hands, not both contradicts dexterity: simultaneously covering 22-DOF five-finger hands and standard two-finger grippers", "strength": "strong" }, { "paper_id": "learning_a_thousand_tasks_in_a_day_2026_08", "title": "1,000 Tasks in 1 Day, Relying on Inductive Bias", "url": "https://haiguangboy.com/posts/mt3-thousand-tasks", "relation": "same_track", "summary": "A point masked by the numbers: there remains a huge gap between 66.7% on dexterous tasks and 90%+ on basic tasks validates perception dependence: vision-only, single camera, no tactile sensing, relying on accurate segmentation", "strength": "strong" }, { "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": "same_track", "summary": "Three-stage recipe: simulation pretraining → train discriminator → weighted real-robot post-training validates ablation confirming the three components contribute independently and additively (pretraining/post-training/RTC)", "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 Sensing Needs Separate Modeling", "url": "https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation", "relation": "same_track", "summary": "T-Rex: Why Tactile Sensing Needs Separate Modeling", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "The most noteworthy element is the discriminator. It requires no manual quality annotation—it uses an already-trained policy as a ruler to measure how much each demonstration resembles actions it recognizes, then converts that score into training weights. This approach is reusable for any system relying on human-collected data, and it is isomorphic to another judgment: when data is scarce, don't let one model learn multiple things at once; offload whatever can be offloaded, and quality filtering counts as one such thing.\n\nAnother point is a divergence. This paper concludes that simulation can only guide basic skills, and dexterity must rely on real-robot data—pure simulation is near zero on dexterous tasks. The other route bets the opposite: pretraining uses almost no real-robot data, and after deployment, failure retry plus a small amount of real-robot RL fills the gap. Both sides have empirical evidence; the bet is on whether dexterity can be simulated at all, and the direction should become clear within six months." }, "ruling": { "importance_score": 3.0, "one_sentence": "The noise in teleoperation data must enter the loss function" }, "asset_base_url": "https://static.haiguangboy.com/papers/dexora", "assets": [ { "type": "pdf_screenshot", "object_key": "papers/dexora/page_01.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/dexora/page_01.webp", "role": "paper_first_page", "size_bytes": 269282 }, { "type": "pdf_screenshot", "object_key": "papers/dexora/key_figure.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/dexora/key_figure.webp", "role": "method_figure", "size_bytes": 207664 }, { "type": "cover_image", "object_key": "papers/dexora/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/dexora/cover.webp", "role": "post_cover", "size_bytes": 81186 }, { "type": "public_brief", "object_key": "papers/dexora/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/dexora/public_brief.md", "role": "public_brief", "size_bytes": 4442 }, { "type": "public_manifest", "object_key": "papers/dexora/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/dexora/public_manifest.json", "role": "public_manifest", "size_bytes": 6588 } ], "published_at": "2026-08-05T18:19:52+08:00", "created_at": "2026-08-05T18:19:52+08:00", "updated_at": "2026-09-02T10:55:06+08:00" }