{"id":"bb171070796779f2c486038f38c73eec0c9d3231a92294654f537078718d9fb5","title":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman argue in Anthropic that robot access matters","argument":"Anthropic published this research on July 9, 2026. Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman report tests of language models controlling simulated and physical robots. They find that performance depends heavily on how a model is connected to a robot: models mostly fail when directly controlling the joints, but can complete some real navigation and manipulation tasks when supervising a pretrained controller or using orientation tools.\n\nThe authors argue that this matters for safety because a model can affect the physical world without controlling a robot’s joints itself; access to a capable controller may be enough. In their tests, pretrained policies raised task performance, but models still struggled with spatial memory, long plans and reliable task completion. In one physical-robot run, a model misjudged a trash can as out of the way; the robot struck and dragged it before the team stopped it.\n\nThe authors conclude that evaluations and deployments should treat access level as part of the system. They call for clearer limits on physical access, so a system can affect some objects while being blocked from others. Real-world trials were limited, and how well the results carry over to other robots, tasks and access controls remains unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published July 9, 2026. The authors describe limited real-world trials; those results do not establish how performance generalizes beyond the tested robots and tasks. The reported results concern the authors’ test suite and selected control interfaces; they do not establish performance for every robot or deployment.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"How Claude performs on robotics tasks","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/claude-plays-robotics","publishedAt":"2026-07-09T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors report that a model’s robotics capability depends substantially on its control interface; direct joint control mostly fails, while higher-level assistance can support some real tasks.","excerpt":"When they must drive the joints themselves they mostly fail. But when they supervise a pretrained controller or use simple orientation tools, they can complete real navigation and manipulation tasks.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors say access to a capable controller matters for safety because a model can affect the physical world without controlling robot joints directly.","excerpt":"a model does not need to drive joints itself to act capably in the world, only access to a competent controller.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"In a physical-robot run, the authors say the model misjudged a trash can and the robot struck and dragged it before the team stopped it.","excerpt":"The trash can was in fact directly in front of the dog. It walked into it, got a leg caught, and dragged the can for a couple of meters before we stopped it.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors conclude that evaluations and deployments should account for system access, and call for clearer limits on physical access.","excerpt":"Evaluations and deployments need to treat access level as a core part of the system, because small changes in tools or control can produce large changes in capability.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors say real-world explorations were limited in trial count and generally aligned with simulation while revealing failure cases.","excerpt":"Although we could not achieve high N count trials due to the serial nature of work in the real world, our explorations generally aligned with our simulated findings","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well would these findings carry over to robots, tasks and access controls beyond those tested?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman argue in Anthropic that robot access matters","summary":"Anthropic published this research on July 9, 2026. Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman report tests of language models controlling simulated and physical robots. They find that performance depends heavily on how a model is connected to a robot: models mostly fail when directly controlling the joints, but can complete some real navigation and manipulation tasks when supervising a pretrained controller or using orientation tools.\n\nThe authors argue that this matters for safety because a model can affect the physical world without controlling a robot’s joints itself; access to a capable controller may be enough. In their tests, pretrained policies raised task performance, but models still struggled with spatial memory, long plans and reliable task completion. In one physical-robot run, a model misjudged a trash can as out of the way; the robot struck and dragged it before the team stopped it.\n\nThe authors conclude that evaluations and deployments should treat access level as part of the system. They call for clearer limits on physical access, so a system can affect some objects while being blocked from others. Real-world trials were limited, and how well the results carry over to other robots, tasks and access controls remains unknown.","question":"How well would these findings carry over to robots, tasks and access controls beyond those tested?","claims":[{"text":"The authors report that a model’s robotics capability depends substantially on its control interface; direct joint control mostly fails, while higher-level assistance can support some real tasks.","excerpt":"When they must drive the joints themselves they mostly fail. But when they supervise a pretrained controller or use simple orientation tools, they can complete real navigation and manipulation tasks.","support":"source-stated"},{"text":"The authors say access to a capable controller matters for safety because a model can affect the physical world without controlling robot joints directly.","excerpt":"a model does not need to drive joints itself to act capably in the world, only access to a competent controller.","support":"source-stated"},{"text":"In a physical-robot run, the authors say the model misjudged a trash can and the robot struck and dragged it before the team stopped it.","excerpt":"The trash can was in fact directly in front of the dog. It walked into it, got a leg caught, and dragged the can for a couple of meters before we stopped it.","support":"source-stated"},{"text":"The authors conclude that evaluations and deployments should account for system access, and call for clearer limits on physical access.","excerpt":"Evaluations and deployments need to treat access level as a core part of the system, because small changes in tools or control can produce large changes in capability.","support":"source-stated"},{"text":"The authors say real-world explorations were limited in trial count and generally aligned with simulation while revealing failure cases.","excerpt":"Although we could not achieve high N count trials due to the serial nature of work in the real world, our explorations generally aligned with our simulated findings","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published July 9, 2026.","The authors describe limited real-world trials; those results do not establish how performance generalizes beyond the tested robots and tasks.","The reported results concern the authors’ test suite and selected control interfaces; they do not establish performance for every robot or deployment."],"mediaId":null,"correctionReason":null,"id":"bb171070796779f2c486038f38c73eec0c9d3231a92294654f537078718d9fb5","publicationId":"c5fcfad6-92c4-9ba7-c6dc-167392f0f84e","version":1,"sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","sourceTitle":"How Claude performs on robotics tasks","publisher":"Anthropic","publishedAt":"2026-07-09T00:00:00.000Z","originalPublished":{"value":"2026-07-09T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-20T02:47:20.265Z","createdAt":"2026-11-07T04:33:24.959Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"1a702219b7ba8f9ccb37aef79b5234ac1d5c88d4d591f96c8214d7b671434544","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.62,"decidedAt":"2026-10-27T21:35:28.689Z"},"provenance":{"textHash":"492785ba29a8caf5643d7958d25db11a940c7459885a11bba2d8de77b7e5729c","observationId":"0e3293813992a995d4c6f5baa718f5e442423b6bd046ffdacd6e27a9ab1e52d3","retrievedAt":"2026-10-27T21:37:20.343Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"0c76e16d-0fbf-f8c4-ef72-bc88e0bc7a77"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"dbffe413-e0ab-3047-855b-ef77afd927d6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-07T04:30:23.649Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/shmuel-berman-michael-ilie-jia-deng-and-daniel-freeman-argue-in-bb171070","formats":{"html":"/readings/shmuel-berman-michael-ilie-jia-deng-and-daniel-freeman-argue-in-bb171070","markdown":"/readings/shmuel-berman-michael-ilie-jia-deng-and-daniel-freeman-argue-in-bb171070.md","json":"/readings/shmuel-berman-michael-ilie-jia-deng-and-daniel-freeman-argue-in-bb171070.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-pre-training","name":"Pre-training","explanation":"A model's initial training on a large dataset, before further adaptation for particular uses.","referenceUrl":"https://developers.google.com/machine-learning/glossary#pre-training"}],"entities":[{"id":"anthropic","name":"Anthropic","kind":"organization","context":"Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.","referenceUrls":["https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/anthropic","json":"/entities/anthropic.json","markdown":"/entities/anthropic.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"bb171070796779f2c486038f38c73eec0c9d3231a92294654f537078718d9fb5","revision":"c5fcfad6-92c4-9ba7-c6dc-167392f0f84e","revisionKind":"publication-id","title":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman argue in Anthropic that robot access matters","url":"https://escapemeridian.com/readings/shmuel-berman-michael-ilie-jia-deng-and-daniel-freeman-argue-in-bb171070"},"source":{"type":"source","id":"https://www.anthropic.com/research/claude-plays-robotics","revision":"492785ba29a8caf5643d7958d25db11a940c7459885a11bba2d8de77b7e5729c","revisionKind":"sha256-source-text","title":"How Claude performs on robotics tasks","url":"https://www.anthropic.com/research/claude-plays-robotics"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"bb171070796779f2c486038f38c73eec0c9d3231a92294654f537078718d9fb5","title":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman argue in Anthropic that robot access matters","argument":"Anthropic published this research on July 9, 2026. Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman report tests of language models controlling simulated and physical robots. They find that performance depends heavily on how a model is connected to a robot: models mostly fail when directly controlling the joints, but can complete some real navigation and manipulation tasks when supervising a pretrained controller or using orientation tools.\n\nThe authors argue that this matters for safety because a model can affect the physical world without controlling a robot’s joints itself; access to a capable controller may be enough. In their tests, pretrained policies raised task performance, but models still struggled with spatial memory, long plans and reliable task completion. In one physical-robot run, a model misjudged a trash can as out of the way; the robot struck and dragged it before the team stopped it.\n\nThe authors conclude that evaluations and deployments should treat access level as part of the system. They call for clearer limits on physical access, so a system can affect some objects while being blocked from others. Real-world trials were limited, and how well the results carry over to other robots, tasks and access controls remains unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published July 9, 2026. The authors describe limited real-world trials; those results do not establish how performance generalizes beyond the tested robots and tasks. The reported results concern the authors’ test suite and selected control interfaces; they do not establish performance for every robot or deployment.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"How Claude performs on robotics tasks","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/claude-plays-robotics","publishedAt":"2026-07-09T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors report that a model’s robotics capability depends substantially on its control interface; direct joint control mostly fails, while higher-level assistance can support some real tasks.","excerpt":"When they must drive the joints themselves they mostly fail. But when they supervise a pretrained controller or use simple orientation tools, they can complete real navigation and manipulation tasks.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors say access to a capable controller matters for safety because a model can affect the physical world without controlling robot joints directly.","excerpt":"a model does not need to drive joints itself to act capably in the world, only access to a competent controller.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"In a physical-robot run, the authors say the model misjudged a trash can and the robot struck and dragged it before the team stopped it.","excerpt":"The trash can was in fact directly in front of the dog. It walked into it, got a leg caught, and dragged the can for a couple of meters before we stopped it.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors conclude that evaluations and deployments should account for system access, and call for clearer limits on physical access.","excerpt":"Evaluations and deployments need to treat access level as a core part of the system, because small changes in tools or control can produce large changes in capability.","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"},{"text":"The authors say real-world explorations were limited in trial count and generally aligned with simulation while revealing failure cases.","excerpt":"Although we could not achieve high N count trials due to the serial nature of work in the real world, our explorations generally aligned with our simulated findings","url":"https://www.anthropic.com/research/claude-plays-robotics","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well would these findings carry over to robots, tasks and access controls beyond those tested?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman argue in Anthropic that robot access matters","summary":"Anthropic published this research on July 9, 2026. Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman report tests of language models controlling simulated and physical robots. They find that performance depends heavily on how a model is connected to a robot: models mostly fail when directly controlling the joints, but can complete some real navigation and manipulation tasks when supervising a pretrained controller or using orientation tools.\n\nThe authors argue that this matters for safety because a model can affect the physical world without controlling a robot’s joints itself; access to a capable controller may be enough. In their tests, pretrained policies raised task performance, but models still struggled with spatial memory, long plans and reliable task completion. In one physical-robot run, a model misjudged a trash can as out of the way; the robot struck and dragged it before the team stopped it.\n\nThe authors conclude that evaluations and deployments should treat access level as part of the system. They call for clearer limits on physical access, so a system can affect some objects while being blocked from others. Real-world trials were limited, and how well the results carry over to other robots, tasks and access controls remains unknown.","question":"How well would these findings carry over to robots, tasks and access controls beyond those tested?","claims":[{"text":"The authors report that a model’s robotics capability depends substantially on its control interface; direct joint control mostly fails, while higher-level assistance can support some real tasks.","excerpt":"When they must drive the joints themselves they mostly fail. But when they supervise a pretrained controller or use simple orientation tools, they can complete real navigation and manipulation tasks.","support":"source-stated"},{"text":"The authors say access to a capable controller matters for safety because a model can affect the physical world without controlling robot joints directly.","excerpt":"a model does not need to drive joints itself to act capably in the world, only access to a competent controller.","support":"source-stated"},{"text":"In a physical-robot run, the authors say the model misjudged a trash can and the robot struck and dragged it before the team stopped it.","excerpt":"The trash can was in fact directly in front of the dog. It walked into it, got a leg caught, and dragged the can for a couple of meters before we stopped it.","support":"source-stated"},{"text":"The authors conclude that evaluations and deployments should account for system access, and call for clearer limits on physical access.","excerpt":"Evaluations and deployments need to treat access level as a core part of the system, because small changes in tools or control can produce large changes in capability.","support":"source-stated"},{"text":"The authors say real-world explorations were limited in trial count and generally aligned with simulation while revealing failure cases.","excerpt":"Although we could not achieve high N count trials due to the serial nature of work in the real world, our explorations generally aligned with our simulated findings","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published July 9, 2026.","The authors describe limited real-world trials; those results do not establish how performance generalizes beyond the tested robots and tasks.","The reported results concern the authors’ test suite and selected control interfaces; they do not establish performance for every robot or deployment."],"mediaId":null,"correctionReason":null,"id":"bb171070796779f2c486038f38c73eec0c9d3231a92294654f537078718d9fb5","publicationId":"c5fcfad6-92c4-9ba7-c6dc-167392f0f84e","version":1,"sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","sourceTitle":"How Claude performs on robotics tasks","publisher":"Anthropic","publishedAt":"2026-07-09T00:00:00.000Z","originalPublished":{"value":"2026-07-09T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-20T02:47:20.265Z","createdAt":"2026-11-07T04:33:24.959Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/claude-plays-robotics","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"1a702219b7ba8f9ccb37aef79b5234ac1d5c88d4d591f96c8214d7b671434544","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.62,"decidedAt":"2026-10-27T21:35:28.689Z"},"provenance":{"textHash":"492785ba29a8caf5643d7958d25db11a940c7459885a11bba2d8de77b7e5729c","observationId":"0e3293813992a995d4c6f5baa718f5e442423b6bd046ffdacd6e27a9ab1e52d3","retrievedAt":"2026-10-27T21:37:20.343Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"0c76e16d-0fbf-f8c4-ef72-bc88e0bc7a77"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"dbffe413-e0ab-3047-855b-ef77afd927d6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-07T04:30:23.649Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"c5fcfad6-92c4-9ba7-c6dc-167392f0f84e","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly attributes the research to Shmuel Berman, Michael Ilie, Jia Deng and Daniel Freeman, published by Anthropic. Findings and safety conclusions are framed as the authors' reports and arguments, not as established fact. The title's phrase 'argue in Anthropic' is awkward but does not misattribute.","excerpt":"Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman"},{"dimension":"summary","status":"uncertain","explanation":"The summary accurately conveys the interface-dependence finding and the safety argument. It drops the source's other central thesis: models are improving quickly across generations, and the reliability gap is 'closing with each model generation.' Without it, the summary reads as mostly limitations and loses the authors' force. Smallest repair: add one attributed sentence saying the authors report rapid, uneven generational gains, especially with higher-level control. 'Policy' and 'orientation tools' are also unexplained beyond the pretrained definition. A brief gloss, such as a separate network turning high-level commands into joint movements, would help.","excerpt":"Models are getting better at robotics quickly, but we found that how capable they are depends heavily on how they are connected to the robot"},{"dimension":"date","status":"pass","explanation":"The reading states July 9, 2026, which matches the UTC calendar day of evidence.originalPublished, and labels the text as retrieved on October 27, 2026. The page was modified on October 23, 2026, but the reading relies on core body passages that appear integral to the original article. There is no hindsight framing, and the authors' expectations are not presented as outcomes. The evidence is a publisher page, not an archive copy.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Each claim matches its quoted source text. That includes the interface dependence, the controller-access safety point, the trash-can incident, the conclusion on access level and the limited real-world trials. The statement that access to a capable controller 'may be enough' stays within the source's wording. The open question about generalization follows from the authors' own stated limits. No benefits or costs are invented.","excerpt":"a model does not need to drive joints itself to act capably in the world, only access to a competent controller."},{"dimension":"media","status":"pass","explanation":"No media was selected, and no verified media candidates were supplied, so nothing is misidentified or invented.","excerpt":null}],"createdAt":"2026-11-07T04:30:23.649Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}