{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:PCNMERUM4KEKZOQI4VMJ7NJHOS","short_pith_number":"pith:PCNMERUM","schema_version":"1.0","canonical_sha256":"789ac2468ce288acba08e5589fb527749ebe5668847bb47664ccf41ac702ca3a","source":{"kind":"arxiv","id":"2605.03065","version":2},"attestation_state":"computed","paper":{"title":"OGPO: Sample Efficient Full-Finetuning of Generative Control Policies","license":"http://creativecommons.org/licenses/by/4.0/","headline":"OGPO fine-tunes generative control policies to near full task success from poor initializations with no expert data.","cross_cats":["cs.RO"],"primary_cat":"cs.LG","authors_text":"Abhishek Gupta, Chaoyi Pan, Cleah Winston, Douglas Chen, Giri Anantharaman, Hongkai Dai, Jesse Zhang, Manan Agarwal, Max Simchowitz, Mitsuhiko Nakamoto, Nai-Chieh Huang, Oliver Kroemer, Paarth Shah, Sarvesh Patil, Sergey Levine, Shashwat Saxena, Zeynep Temel","submitted_at":"2026-05-04T18:36:40Z","abstract_excerpt":"Generative control policies (GCPs), such as diffusion- and flow-based control policies, have emerged as effective parameterizations for robot learning. This work introduces Off-policy Generative Policy Optimization (OGPO), a sample-efficient algorithm for finetuning GCPs that maintains off-policy critic networks to maximize data reuse and propagate policy gradients through the full generative process of the policy via a modified PPO objective, using critics as the terminal reward. OGPO achieves state-of-the-art performance on manipulation tasks spanning multi-task settings, high-precision inse"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.03065","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T18:36:40Z","cross_cats_sorted":["cs.RO"],"title_canon_sha256":"8e90aa034719a618774f63ed25aaa69d7d1dba226830d5d96d7cdf3d53126fe5","abstract_canon_sha256":"3a1340d177da08657e0d0a1971623f52c995294951d18469f33b5163c4185185"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-11T01:09:36.935088Z","signature_b64":"smqEyff/Vj6HZTYL4/sI02ejykVkACbB7VeAUh0n5dfblccdQIFF7TaDniIkoN8dh+tSl7jvtuz0qbp+/bF7Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"789ac2468ce288acba08e5589fb527749ebe5668847bb47664ccf41ac702ca3a","last_reissued_at":"2026-06-11T01:09:36.934198Z","signature_status":"signed_v1","first_computed_at":"2026-06-11T01:09:36.934198Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"OGPO: Sample Efficient Full-Finetuning of Generative Control Policies","license":"http://creativecommons.org/licenses/by/4.0/","headline":"OGPO fine-tunes generative control policies to near full task success from poor initializations with no expert data.","cross_cats":["cs.RO"],"primary_cat":"cs.LG","authors_text":"Abhishek Gupta, Chaoyi Pan, Cleah Winston, Douglas Chen, Giri Anantharaman, Hongkai Dai, Jesse Zhang, Manan Agarwal, Max Simchowitz, Mitsuhiko Nakamoto, Nai-Chieh Huang, Oliver Kroemer, Paarth Shah, Sarvesh Patil, Sergey Levine, Shashwat Saxena, Zeynep Temel","submitted_at":"2026-05-04T18:36:40Z","abstract_excerpt":"Generative control policies (GCPs), such as diffusion- and flow-based control policies, have emerged as effective parameterizations for robot learning. This work introduces Off-policy Generative Policy Optimization (OGPO), a sample-efficient algorithm for finetuning GCPs that maintains off-policy critic networks to maximize data reuse and propagate policy gradients through the full generative process of the policy via a modified PPO objective, using critics as the terminal reward. OGPO achieves state-of-the-art performance on manipulation tasks spanning multi-task settings, high-precision inse"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"OGPO achieves state-of-the-art performance on manipulation tasks spanning multi-task settings, high-precision insertion, and dexterous control. It is the only method that can fine-tune poorly-initialized behavior cloning policies to near full task-success with no expert data in the online replay buffer, and does so with few task-specific hyperparameter tuning.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That off-policy critics can be maintained stably while propagating gradients through the full generative process of the policy via a modified PPO objective without introducing instability or requiring heavy per-task tuning across state- and pixel-based settings.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OGPO is a sample-efficient off-policy method for full finetuning of generative control policies that reaches SOTA on robotic manipulation tasks and can recover from poor behavior-cloning initializations without expert data.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"OGPO fine-tunes generative control policies to near full task success from poor initializations with no expert data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2c21246922662e0d42dda07e7ca12555ba657f22bf3cf7eba0abc5d90f06eb2f"},"source":{"id":"2605.03065","kind":"arxiv","version":2},"verdict":{"id":"e5507149-99c4-4129-b69a-8b673fed22ce","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T18:41:53.828658Z","strongest_claim":"OGPO achieves state-of-the-art performance on manipulation tasks spanning multi-task settings, high-precision insertion, and dexterous control. It is the only method that can fine-tune poorly-initialized behavior cloning policies to near full task-success with no expert data in the online replay buffer, and does so with few task-specific hyperparameter tuning.","one_line_summary":"OGPO is a sample-efficient off-policy method for full finetuning of generative control policies that reaches SOTA on robotic manipulation tasks and can recover from poor behavior-cloning initializations without expert data.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That off-policy critics can be maintained stably while propagating gradients through the full generative process of the policy via a modified PPO objective without introducing instability or requiring heavy per-task tuning across state- and pixel-based settings.","pith_extraction_headline":"OGPO fine-tunes generative control policies to near full task success from poor initializations with no expert data."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.03065/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:38:37.362641Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T02:01:22.017696Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:44:35.772770Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"79a51aaa64822ad9c6ff6402b34c61edc00301dc603d44c675650b897c858442"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"4b1db2abbe20e034c83416194efca1c0cfd7d40e207ec3a70c921a05b577b489"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.03065","created_at":"2026-06-11T01:09:36.934308+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.03065v2","created_at":"2026-06-11T01:09:36.934308+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03065","created_at":"2026-06-11T01:09:36.934308+00:00"},{"alias_kind":"pith_short_12","alias_value":"PCNMERUM4KEK","created_at":"2026-06-11T01:09:36.934308+00:00"},{"alias_kind":"pith_short_16","alias_value":"PCNMERUM4KEKZOQI","created_at":"2026-06-11T01:09:36.934308+00:00"},{"alias_kind":"pith_short_8","alias_value":"PCNMERUM","created_at":"2026-06-11T01:09:36.934308+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS","json":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS.json","graph_json":"https://pith.science/api/pith-number/PCNMERUM4KEKZOQI4VMJ7NJHOS/graph.json","events_json":"https://pith.science/api/pith-number/PCNMERUM4KEKZOQI4VMJ7NJHOS/events.json","paper":"https://pith.science/paper/PCNMERUM"},"agent_actions":{"view_html":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS","download_json":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS.json","view_paper":"https://pith.science/paper/PCNMERUM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.03065&json=true","fetch_graph":"https://pith.science/api/pith-number/PCNMERUM4KEKZOQI4VMJ7NJHOS/graph.json","fetch_events":"https://pith.science/api/pith-number/PCNMERUM4KEKZOQI4VMJ7NJHOS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS/action/storage_attestation","attest_author":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS/action/author_attestation","sign_citation":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS/action/citation_signature","submit_replication":"https://pith.science/pith/PCNMERUM4KEKZOQI4VMJ7NJHOS/action/replication_record"}},"created_at":"2026-06-11T01:09:36.934308+00:00","updated_at":"2026-06-11T01:09:36.934308+00:00"}