{ "schema_version": "paper_public_manifest_v1", "paper_id": "pi-rl-chelsea-finn-talk_2026_07", "slug": "pi-rl-chelsea-finn-talk", "title": "The bottleneck in robot RL is sampling, not algorithms", "authors": [], "source": { "arxiv_id": "", "pdf_url": "", "project_url": "https://x.com/karolina_dubiel/status/2081534295823589877", "github_url": "", "huggingface_url": "", "original_source": "https://x.com/karolina_dubiel/status/2081534295823589877" }, "site": { "post_url": "/posts/pi-rl-chelsea-finn-talk", "canonical_url": "https://haiguangboy.com/posts/pi-rl-chelsea-finn-talk", "cover_image": "https://static.haiguangboy.com/papers/pi-rl-chelsea-finn-talk/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "world_model", "tasks": [ "embodied_ai", "world_model", "robotics", "Embodied intelligence", "Robot learning", "Reinforcement learning" ], "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": "As reported: general-purpose robot models may already be beating specialized pipelines (per π0.7) contradicts joint co-training on heterogeneous human-robot data being a structurally suboptimal approach", "strength": "strong" }, { "paper_id": "wx_中国财富_20260719_2026_07", "title": "Qianxun Intelligence Han Fengtao: embodied intelligence enters the 'speed competition' phase", "url": "https://haiguangboy.com/posts/qianxun-general-brain-data-loop", "relation": "contrast", "summary": "As reported: the bottleneck for RL in robotics is the cost of physical rollouts, not just algorithm quality contradicts the core judgment: the bottleneck is not in hardware but in the brain; the robot market is only about 1% the size of the auto market", "strength": "strong" }, { "paper_id": "wx_星河频率_20260718_2026_07", "title": "Robots begin to 'stand in the light': Lingchu Intelligence enters the optical module production line", "url": "https://haiguangboy.com/posts/lingchu-optical", "relation": "same_track", "summary": "Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates boundary: everything is Lingchu Intelligence's own claims plus media retelling, with no third-party independent verification of key figures", "strength": "strong" }, { "paper_id": "wx_elsewhere别处发生_20260722_2026_07", "title": "Liang Wenfeng's four-hour investor meeting transcript", "url": "https://haiguangboy.com/posts/liangwenfeng-world-model", "relation": "same_track", "summary": "Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates ★boundary: a compilation of secondhand accounts of a non-public meeting, with the media noting 'some wording may differ slightly from the original,' and no audio or raw records to corroborate", "strength": "strong" }, { "paper_id": "wx_晚点latepost_20260718_2026_07", "title": "Sudu Technology's WAIC debut: reality is the endpoint for robots, not the boundary of training", "url": "https://haiguangboy.com/posts/sudu-third-position", "relation": "same_track", "summary": "Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates boundary: a single founder interview, with core technology and business claims all self-reported, and the 20B valuation/10K-GPU compute/POC progress unverified by third parties", "strength": "strong" }, { "paper_id": "sunday_blog_20260717_2026_07", "title": "sunday_blog_20260717", "url": "https://haiguangboy.com/posts/sunday_blog_20260717", "relation": "same_track", "summary": "As reported: general-purpose robot models may already be beating specialized pipelines (per π0.7) validates the route bet: strong pretraining rewrites the scaling equation—extremely small amounts of internal data drive long-tail improvements", "strength": "medium" } ] }, "ruling": { "importance_score": 3.0, "one_sentence": "The bottleneck in robot RL is sampling, not algorithms" }, "asset_base_url": "https://static.haiguangboy.com/papers/pi-rl-chelsea-finn-talk", "assets": [ { "type": "cover_image", "object_key": "papers/pi-rl-chelsea-finn-talk/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pi-rl-chelsea-finn-talk/cover.webp", "role": "post_cover", "size_bytes": 75094 }, { "type": "public_brief", "object_key": "papers/pi-rl-chelsea-finn-talk/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pi-rl-chelsea-finn-talk/public_brief.md", "role": "public_brief", "size_bytes": 5368 }, { "type": "public_manifest", "object_key": "papers/pi-rl-chelsea-finn-talk/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/pi-rl-chelsea-finn-talk/public_manifest.json", "role": "public_manifest", "size_bytes": 5476 } ], "published_at": "2026-07-29T16:59:36+08:00", "created_at": "2026-07-29T16:59:36+08:00", "updated_at": "2026-09-02T10:55:06+08:00", "analyst_take": { "type": "author_opinion", "text": "Looking at the three judgments in the library side by side, there are three different answers to 'where the bottleneck lies': this one says it's the cost of physical rollouts; another judgment says the bottleneck is not in hardware but in the brain; and one more is even sharper—the implicit premise of 'world models have nothing to do with the ceiling of intelligence' is that the data pipeline is already connected, so for LLMs the world model is a detour, while for robots it's not connected, so it's a bridge.\n\nThe three don't actually conflict. Saying 'the brain' is about the ceiling of capability, while saying 'rollout cost' is about how to realize that capability—they're different segments of the same path. And another judgment in the library—'high-quality references let RL train with an extremely simple formula'—exactly corroborates this reading: the main battlefield is not finding better policy gradient variants, but increasing the information yield per real trajectory.\n\nWe can come back in six months to verify: if subsequent versions of π still don't go end-to-end torque control, it means the structured control interface is not a transitional state but a stable choice." } }