{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:XKO3DYKMVIOKZM5HX6TSVZFV26","short_pith_number":"pith:XKO3DYKM","schema_version":"1.0","canonical_sha256":"ba9db1e14caa1cacb3a7bfa72ae4b5d7ba02910f260d89f95f7fc67d2f8b6f37","source":{"kind":"arxiv","id":"2605.13105","version":1},"attestation_state":"computed","paper":{"title":"What to Ignore, What to React: Visually Robust RL Fine-Tuning of VLA Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants to improve RL fine-tuning of VLA models under visual shifts.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Chuheng Zhang, Jiang Bian, Jingjing Fu, Jun Zhang, Ling Zhang, Li Zhao, Mingyu Liu, Rui Wang, Yuanfang Peng","submitted_at":"2026-05-13T07:15:37Z","abstract_excerpt":"Reinforcement learning (RL) fine-tuning has shown promise for Vision-Language-Action (VLA) models in robotic manipulation, but deployment-time visual shifts pose practical challenges. A key difficulty is that standard task rewards supervise task success, but offer limited guidance on whether a visual change is task-irrelevant or changes the behavior required for manipulation. We propose PAIR-VLA (Paired Action Invariance & Sensitivity for Visually Robust VLA), an RL fine-tuning framework to address this difficulty by adding two auxiliary objectives over paired visual variants during PPO optimi"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13105","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-13T07:15:37Z","cross_cats_sorted":[],"title_canon_sha256":"974139390ca7ed35f9d6b6dc7187f4d9158adabc2396863129e1be71ce3dadfc","abstract_canon_sha256":"2216604d31c8e863c8e4a104c63fcf7c24903ce67f6a7380cbd2b52faf0e1fc4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:08:58.185618Z","signature_b64":"wjt18k2VdFeHoruV1e1jYr7vrUYkyrmVIq/O509yiE0Ek4dNIsPfYtyHvcFteAG80rU8tcqTbs3b8Vglt036CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ba9db1e14caa1cacb3a7bfa72ae4b5d7ba02910f260d89f95f7fc67d2f8b6f37","last_reissued_at":"2026-05-18T03:08:58.185087Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:08:58.185087Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"What to Ignore, What to React: Visually Robust RL Fine-Tuning of VLA Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants to improve RL fine-tuning of VLA models under visual shifts.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Chuheng Zhang, Jiang Bian, Jingjing Fu, Jun Zhang, Ling Zhang, Li Zhao, Mingyu Liu, Rui Wang, Yuanfang Peng","submitted_at":"2026-05-13T07:15:37Z","abstract_excerpt":"Reinforcement learning (RL) fine-tuning has shown promise for Vision-Language-Action (VLA) models in robotic manipulation, but deployment-time visual shifts pose practical challenges. A key difficulty is that standard task rewards supervise task success, but offer limited guidance on whether a visual change is task-irrelevant or changes the behavior required for manipulation. We propose PAIR-VLA (Paired Action Invariance & Sensitivity for Visually Robust VLA), an RL fine-tuning framework to address this difficulty by adding two auxiliary objectives over paired visual variants during PPO optimi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our method consistently improves over standard PPO, achieving average improvements of 16.62% on π0.5 and 9.10% on OpenVLA across diverse out-of-distribution visual shifts.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That paired visual variants (task-preserving and task-altering) can be reliably generated or labeled during training to provide accurate behavior-level supervision without introducing new biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants during PPO fine-tuning of VLA models, yielding 9-16% average gains on ManiSkill3 under distractors, textures, poses, viewpoints, and lighting shifts.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants to improve RL fine-tuning of VLA models under visual shifts.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0d6da9a3407ba9d4c3a6557a00747a9cbcc78446906970bcaaf9a56db4009d45"},"source":{"id":"2605.13105","kind":"arxiv","version":1},"verdict":{"id":"7679d0dc-6da7-487d-b481-ed305e022076","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:41:43.923590Z","strongest_claim":"Our method consistently improves over standard PPO, achieving average improvements of 16.62% on π0.5 and 9.10% on OpenVLA across diverse out-of-distribution visual shifts.","one_line_summary":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants during PPO fine-tuning of VLA models, yielding 9-16% average gains on ManiSkill3 under distractors, textures, poses, viewpoints, and lighting shifts.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That paired visual variants (task-preserving and task-altering) can be reliably generated or labeled during training to provide accurate behavior-level supervision without introducing new biases.","pith_extraction_headline":"PAIR-VLA adds invariance and sensitivity objectives over paired visual variants to improve RL fine-tuning of VLA models under visual shifts."},"references":{"count":43,"sample":[{"doi":"","year":2024,"title":"Open x- embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0","work_id":"846c44cc-0874-4a3c-90cb-b86f68157e99","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset","work_id":"13253de2-3d89-415c-8c2f-3adb25d4c337","ref_index":2,"cited_arxiv_id":"2403.12945","is_internal_anchor":true},{"doi":"","year":2023,"title":"Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023","work_id":"f08c7a53-d673-4be5-b955-d92443959ebf","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Octo: An open-source generalist robot policy, 2024","work_id":"9644ac38-9a21-4d5e-9386-5460ec7456ea","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","ref_index":5,"cited_arxiv_id":"2410.24164","is_internal_anchor":true}],"resolved_work":43,"snapshot_sha256":"ec062a49971027b94d2df6dce5fbe087e484745e564ade7f770ea2cdc4c7f442","internal_anchors":9},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"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":"2605.13105","created_at":"2026-05-18T03:08:58.185161+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13105v1","created_at":"2026-05-18T03:08:58.185161+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13105","created_at":"2026-05-18T03:08:58.185161+00:00"},{"alias_kind":"pith_short_12","alias_value":"XKO3DYKMVIOK","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"XKO3DYKMVIOKZM5H","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"XKO3DYKM","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26","json":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26.json","graph_json":"https://pith.science/api/pith-number/XKO3DYKMVIOKZM5HX6TSVZFV26/graph.json","events_json":"https://pith.science/api/pith-number/XKO3DYKMVIOKZM5HX6TSVZFV26/events.json","paper":"https://pith.science/paper/XKO3DYKM"},"agent_actions":{"view_html":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26","download_json":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26.json","view_paper":"https://pith.science/paper/XKO3DYKM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13105&json=true","fetch_graph":"https://pith.science/api/pith-number/XKO3DYKMVIOKZM5HX6TSVZFV26/graph.json","fetch_events":"https://pith.science/api/pith-number/XKO3DYKMVIOKZM5HX6TSVZFV26/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26/action/storage_attestation","attest_author":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26/action/author_attestation","sign_citation":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26/action/citation_signature","submit_replication":"https://pith.science/pith/XKO3DYKMVIOKZM5HX6TSVZFV26/action/replication_record"}},"created_at":"2026-05-18T03:08:58.185161+00:00","updated_at":"2026-05-18T03:08:58.185161+00:00"}