{"id":"2e1c55f406d792274297e000491229043f41b0b8cf2881c15947913828a5dfb8","title":"Yuxiang Zhou and co-authors report in Google DeepMind: RoboCat for robotic manipulation","argument":"The authors’ publication is dated December 5, 2023. They propose RoboCat, a foundation agent for robotic manipulation that learns from visual experience labelled with actions across different robots and tasks. They report that it can generalise to new tasks and robots zero-shot, or adapt using 100–1000 examples for a target task. The authors also use a trained model to generate data for later training iterations, describing this as a basic building block for an autonomous improvement loop. In evaluations in simulation and on three real robot embodiments, they report that growing and diversifying training data brings signs of cross-task transfer and makes RoboCat more efficient at adapting to new tasks. The stated significance is that experience across robots and tasks could help robots acquire novel skills and embodiments, though the abstract does not establish how well these results carry over to broader real-world use.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 2, 2026; originally published December 5, 2023. The abstract does not specify the tasks, evaluation measures, or comparison baseline behind the reported transfer and adaptation results. The abstract does not establish how the reported results extend to broader real-world use.","speaker":{"name":"Google DeepMind","roleAtSource":"Publisher"},"source":{"title":"RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation","publisher":"Google DeepMind","format":"Article","url":"https://deepmind.google/research/publications/35829/","publishedAt":"2023-12-05T00:00:00+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors propose RoboCat as a foundation agent for robotic manipulation, using visual goal-conditioned decision transformer architecture and experience from multiple robots and tasks.","excerpt":"we propose a foundation agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"The authors report generalisation to new tasks and robots, both zero-shot and through adaptation using 100–1000 examples for the target task.","excerpt":"generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100--1000 examples for the target task","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"The authors report using a trained model to generate data for subsequent training iterations as a basic building block for an autonomous improvement loop.","excerpt":"a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"In evaluations in simulation and on three real robot embodiments, the authors report signs of cross-task transfer and greater efficiency adapting to new tasks as training data grows and diversifies.","excerpt":"as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/35829/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does RoboCat’s reported adaptation and cross-task transfer extend to real-world tasks beyond the evaluations described in the abstract?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Yuxiang Zhou and co-authors report in Google DeepMind: RoboCat for robotic manipulation","summary":"The authors’ publication is dated December 5, 2023. They propose RoboCat, a foundation agent for robotic manipulation that learns from visual experience labelled with actions across different robots and tasks. They report that it can generalise to new tasks and robots zero-shot, or adapt using 100–1000 examples for a target task. The authors also use a trained model to generate data for later training iterations, describing this as a basic building block for an autonomous improvement loop. In evaluations in simulation and on three real robot embodiments, they report that growing and diversifying training data brings signs of cross-task transfer and makes RoboCat more efficient at adapting to new tasks. The stated significance is that experience across robots and tasks could help robots acquire novel skills and embodiments, though the abstract does not establish how well these results carry over to broader real-world use.","question":"How well does RoboCat’s reported adaptation and cross-task transfer extend to real-world tasks beyond the evaluations described in the abstract?","claims":[{"text":"The authors propose RoboCat as a foundation agent for robotic manipulation, using visual goal-conditioned decision transformer architecture and experience from multiple robots and tasks.","excerpt":"we propose a foundation agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer","support":"source-stated"},{"text":"The authors report generalisation to new tasks and robots, both zero-shot and through adaptation using 100–1000 examples for the target task.","excerpt":"generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100--1000 examples for the target task","support":"source-stated"},{"text":"The authors report using a trained model to generate data for subsequent training iterations as a basic building block for an autonomous improvement loop.","excerpt":"a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop","support":"source-stated"},{"text":"In evaluations in simulation and on three real robot embodiments, the authors report signs of cross-task transfer and greater efficiency adapting to new tasks as training data grows and diversifies.","excerpt":"as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.","support":"source-stated"}],"limitations":["Text as retrieved on November 2, 2026; originally published December 5, 2023.","The abstract does not specify the tasks, evaluation measures, or comparison baseline behind the reported transfer and adaptation results.","The abstract does not establish how the reported results extend to broader real-world use."],"mediaId":null,"correctionReason":null,"id":"2e1c55f406d792274297e000491229043f41b0b8cf2881c15947913828a5dfb8","publicationId":"dac24c8f-5ce9-d13e-5479-c649c7167881","version":1,"sourceUrl":"https://deepmind.google/research/publications/35829/","sourceTitle":"RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation","publisher":"Google DeepMind","publishedAt":"2023-12-05T00:00:00+00:00","originalPublished":{"value":"2023-12-05T00:00:00+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-11-02T16:15:45.301Z","createdAt":"2026-11-22T09:38:19.637Z","historical":true,"entityIds":["google-deepmind"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/35829/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"7ad447370968103eb7e7532d2138c7943ed1bdb8f4adbe674fd48db9986be0aa","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-02T16:15:43.388Z"},"provenance":{"textHash":"d41d4ffee623a42bb21e64a9cb4c1bccafca750b105fb7301b2368820403efd6","observationId":"e521a67f8415db5d9ec9b78e8ce7562a543de107f8b939bae9673a2ec0dda172","retrievedAt":"2026-11-02T16:17:23.078Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"133b3429-ce4f-ea2e-4553-4fd197de3807"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"229efa32-ac3b-0bfc-4da1-64631cba8bee"}},"audit":{"status":"passed","reviewedAt":"2026-11-22T09:33:29.766Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/yuxiang-zhou-and-co-authors-report-in-google-deepmind-robocat-2e1c55f4","formats":{"html":"/readings/yuxiang-zhou-and-co-authors-report-in-google-deepmind-robocat-2e1c55f4","markdown":"/readings/yuxiang-zhou-and-co-authors-report-in-google-deepmind-robocat-2e1c55f4.md","json":"/readings/yuxiang-zhou-and-co-authors-report-in-google-deepmind-robocat-2e1c55f4.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":[],"conditions":[],"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-zero-shot","name":"Zero-shot","explanation":"Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).","referenceUrl":"https://developers.google.com/machine-learning/glossary#zero-shot-prompting"}],"entities":[{"id":"google-deepmind","name":"Google DeepMind","kind":"organization","context":"Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.","referenceUrls":["https://deepmind.google/blog/","https://deepmind.google/science/alphafold/"],"path":"/entities/google-deepmind","json":"/entities/google-deepmind.json","markdown":"/entities/google-deepmind.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"2e1c55f406d792274297e000491229043f41b0b8cf2881c15947913828a5dfb8","revision":"dac24c8f-5ce9-d13e-5479-c649c7167881","revisionKind":"publication-id","title":"Yuxiang Zhou and co-authors report in Google DeepMind: RoboCat for robotic manipulation","url":"https://escapemeridian.com/readings/yuxiang-zhou-and-co-authors-report-in-google-deepmind-robocat-2e1c55f4"},"source":{"type":"source","id":"https://deepmind.google/research/publications/35829/","revision":"d41d4ffee623a42bb21e64a9cb4c1bccafca750b105fb7301b2368820403efd6","revisionKind":"sha256-source-text","title":"RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation","url":"https://deepmind.google/research/publications/35829/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"2e1c55f406d792274297e000491229043f41b0b8cf2881c15947913828a5dfb8","title":"Yuxiang Zhou and co-authors report in Google DeepMind: RoboCat for robotic manipulation","argument":"The authors’ publication is dated December 5, 2023. They propose RoboCat, a foundation agent for robotic manipulation that learns from visual experience labelled with actions across different robots and tasks. They report that it can generalise to new tasks and robots zero-shot, or adapt using 100–1000 examples for a target task. The authors also use a trained model to generate data for later training iterations, describing this as a basic building block for an autonomous improvement loop. In evaluations in simulation and on three real robot embodiments, they report that growing and diversifying training data brings signs of cross-task transfer and makes RoboCat more efficient at adapting to new tasks. The stated significance is that experience across robots and tasks could help robots acquire novel skills and embodiments, though the abstract does not establish how well these results carry over to broader real-world use.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 2, 2026; originally published December 5, 2023. The abstract does not specify the tasks, evaluation measures, or comparison baseline behind the reported transfer and adaptation results. The abstract does not establish how the reported results extend to broader real-world use.","speaker":{"name":"Google DeepMind","roleAtSource":"Publisher"},"source":{"title":"RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation","publisher":"Google DeepMind","format":"Article","url":"https://deepmind.google/research/publications/35829/","publishedAt":"2023-12-05T00:00:00+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"The authors propose RoboCat as a foundation agent for robotic manipulation, using visual goal-conditioned decision transformer architecture and experience from multiple robots and tasks.","excerpt":"we propose a foundation agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"The authors report generalisation to new tasks and robots, both zero-shot and through adaptation using 100–1000 examples for the target task.","excerpt":"generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100--1000 examples for the target task","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"The authors report using a trained model to generate data for subsequent training iterations as a basic building block for an autonomous improvement loop.","excerpt":"a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"},{"text":"In evaluations in simulation and on three real robot embodiments, the authors report signs of cross-task transfer and greater efficiency adapting to new tasks as training data grows and diversifies.","excerpt":"as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.","url":"https://deepmind.google/research/publications/35829/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/35829/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does RoboCat’s reported adaptation and cross-task transfer extend to real-world tasks beyond the evaluations described in the abstract?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Yuxiang Zhou and co-authors report in Google DeepMind: RoboCat for robotic manipulation","summary":"The authors’ publication is dated December 5, 2023. They propose RoboCat, a foundation agent for robotic manipulation that learns from visual experience labelled with actions across different robots and tasks. They report that it can generalise to new tasks and robots zero-shot, or adapt using 100–1000 examples for a target task. The authors also use a trained model to generate data for later training iterations, describing this as a basic building block for an autonomous improvement loop. In evaluations in simulation and on three real robot embodiments, they report that growing and diversifying training data brings signs of cross-task transfer and makes RoboCat more efficient at adapting to new tasks. The stated significance is that experience across robots and tasks could help robots acquire novel skills and embodiments, though the abstract does not establish how well these results carry over to broader real-world use.","question":"How well does RoboCat’s reported adaptation and cross-task transfer extend to real-world tasks beyond the evaluations described in the abstract?","claims":[{"text":"The authors propose RoboCat as a foundation agent for robotic manipulation, using visual goal-conditioned decision transformer architecture and experience from multiple robots and tasks.","excerpt":"we propose a foundation agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer","support":"source-stated"},{"text":"The authors report generalisation to new tasks and robots, both zero-shot and through adaptation using 100–1000 examples for the target task.","excerpt":"generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100--1000 examples for the target task","support":"source-stated"},{"text":"The authors report using a trained model to generate data for subsequent training iterations as a basic building block for an autonomous improvement loop.","excerpt":"a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop","support":"source-stated"},{"text":"In evaluations in simulation and on three real robot embodiments, the authors report signs of cross-task transfer and greater efficiency adapting to new tasks as training data grows and diversifies.","excerpt":"as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.","support":"source-stated"}],"limitations":["Text as retrieved on November 2, 2026; originally published December 5, 2023.","The abstract does not specify the tasks, evaluation measures, or comparison baseline behind the reported transfer and adaptation results.","The abstract does not establish how the reported results extend to broader real-world use."],"mediaId":null,"correctionReason":null,"id":"2e1c55f406d792274297e000491229043f41b0b8cf2881c15947913828a5dfb8","publicationId":"dac24c8f-5ce9-d13e-5479-c649c7167881","version":1,"sourceUrl":"https://deepmind.google/research/publications/35829/","sourceTitle":"RoboCat: A Self-Improving Foundation Agent for Robotic Manipulation","publisher":"Google DeepMind","publishedAt":"2023-12-05T00:00:00+00:00","originalPublished":{"value":"2023-12-05T00:00:00+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-11-02T16:15:45.301Z","createdAt":"2026-11-22T09:38:19.637Z","historical":true,"entityIds":["google-deepmind"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/35829/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"7ad447370968103eb7e7532d2138c7943ed1bdb8f4adbe674fd48db9986be0aa","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-02T16:15:43.388Z"},"provenance":{"textHash":"d41d4ffee623a42bb21e64a9cb4c1bccafca750b105fb7301b2368820403efd6","observationId":"e521a67f8415db5d9ec9b78e8ce7562a543de107f8b939bae9673a2ec0dda172","retrievedAt":"2026-11-02T16:17:23.078Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"133b3429-ce4f-ea2e-4553-4fd197de3807"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"229efa32-ac3b-0bfc-4da1-64631cba8bee"}},"audit":{"status":"passed","reviewedAt":"2026-11-22T09:33:29.766Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"dac24c8f-5ce9-d13e-5479-c649c7167881","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes RoboCat and its results to the authors listed on the Google DeepMind publication page, and keeps each result as an author report rather than an established fact. The page is a single publication abstract, not a listing.","excerpt":"Yuxiang Zhou, Todor Davchev, Maria Bauza, Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex Lee"},{"dimension":"summary","status":"pass","explanation":"The reading leads with the actor and the proposal. It explains RoboCat in plain terms as an agent that learns from action-labelled visual experience across robots and tasks. It covers the self-generated training data loop, the scaling finding, and the stated potential, and it notes the real-world gap. Zero-shot is defined on the page. Embodiments is understandable from context as different robot bodies. The question follows from the evidence without leading.","excerpt":"has the potential to transform robot learning"},{"dimension":"date","status":"pass","explanation":"The displayed date of December 5, 2023 matches the page dateline and the UTC day of originalPublished, at day precision. The page was modified in 2025, but the cited abstract is plainly the original piece. The reading uses no later events or hindsight framing. The evidence is a publisher page, not an archive copy, so no archive disclosure is needed.","excerpt":"December 5, 2023"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract verbatim in substance: the foundation agent proposal, zero-shot and 100–1000 example adaptation, model-generated training data as a building block, and cross-task transfer with more efficient adaptation as data grows. The significance is correctly framed as potential. The limitations about unspecified baselines and real-world extension are fair. The later page edit does not affect this abstract text.","excerpt":"We find that as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks."},{"dimension":"media","status":"pass","explanation":"No media is selected and no verified media candidates were supplied, so nothing is misidentified or invented.","excerpt":null}],"createdAt":"2026-11-22T09:33:29.766Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}