{"paper":{"title":"Behavior Cloning for Active Perception with Low-Resolution Egocentric Vision","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Behavior cloning from low-resolution egocentric images lets a robot arm actively center a plant for grasping.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Anthony Bilic, Chen Chen, Ladislau B\\\"ol\\\"oni","submitted_at":"2026-05-13T20:45:21Z","abstract_excerpt":"We investigate whether behavior cloning is sufficient to produce active perception in a structured object-finding task. A low-cost robot arm equipped with a wrist-mounted egocentric RGB camera must reposition to center a partially visible plant before triggering a grasp signal, requiring actions that improve future observations. The model predicts joint commands directly from low-resolution RGB images under closed-loop control. We show that low-resolution egocentric vision is sufficient for reliable task completion and that predicting relative joint deltas substantially outperforms absolute jo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that low-resolution egocentric vision is sufficient for reliable task completion and that predicting relative joint deltas substantially outperforms absolute joint position prediction in our setting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That demonstrations collected in the structured plant-finding task provide sufficient coverage for the cloned policy to generalize active perception behaviors to new initial views or slight variations in plant appearance and positioning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Behavior cloning produces active perception in a plant-centering task where a robot arm uses low-resolution egocentric RGB images to predict joint movements, with relative deltas outperforming absolute positions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Behavior cloning from low-resolution egocentric images lets a robot arm actively center a plant for grasping.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"38b8dd87b2d63e30b1d6176159d0c61a620eb9ccddef202d715e29c0e82d9695"},"source":{"id":"2605.14106","kind":"arxiv","version":1},"verdict":{"id":"b6958150-f566-4a28-877d-75eb7b5d4f0a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:05:45.289699Z","strongest_claim":"We show that low-resolution egocentric vision is sufficient for reliable task completion and that predicting relative joint deltas substantially outperforms absolute joint position prediction in our setting.","one_line_summary":"Behavior cloning produces active perception in a plant-centering task where a robot arm uses low-resolution egocentric RGB images to predict joint movements, with relative deltas outperforming absolute positions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That demonstrations collected in the structured plant-finding task provide sufficient coverage for the cloned policy to generalize active perception behaviors to new initial views or slight variations in plant appearance and positioning.","pith_extraction_headline":"Behavior cloning from low-resolution egocentric images lets a robot arm actively center a plant for grasping."},"references":{"count":9,"sample":[{"doi":"","year":2019,"title":"Learning latent dynamics for planning from pixels,","work_id":"d14bee64-f9d5-44cf-8d88-0c1a03405efc","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"An algorithmic perspective on imitation learning,","work_id":"4221725f-7f3a-4c39-b901-2d6ba8807c4c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots","work_id":"1286fc26-002b-4688-a98f-c2e2e6163f9f","ref_index":3,"cited_arxiv_id":"2402.10329","is_internal_anchor":true},{"doi":"","year":2018,"title":"Vision- based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration,","work_id":"337b85c9-eff5-4834-a4ca-f7cca60f455a","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1991,"title":"D. H. Ballard, “Animate vision,”Artificial intelligence, vol. 48, no. 1, pp. 57–86, 1991","work_id":"7e2da00c-bf42-40b8-9694-02b251b23ed1","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":9,"snapshot_sha256":"b04f34e9d548617f23e6309207edc0a996343128a41145d10c9549c640005fd9","internal_anchors":2},"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"}