[
    {
        "id": "osp-16377",
        "type": "article-journal",
        "title": "CAPABLE: Capability-Aware Policy Adaptation via Behavioral Latent Encoding",
        "author": [
            {
                "family": "Khoshnazar",
                "given": "Mohammad"
            },
            {
                "family": "Tezerjani",
                "given": "Mohammad Dehghani"
            },
            {
                "family": "Qu",
                "given": "Deyuan"
            },
            {
                "family": "Gao",
                "given": "Zhiyuan"
            },
            {
                "family": "Zhan",
                "given": "Yanxiang"
            },
            {
                "family": "Schafer",
                "given": "Jeroen"
            },
            {
                "family": "Melnik",
                "given": "Andrew"
            },
            {
                "family": "Yang",
                "given": "Qing"
            },
            {
                "family": "Beetz",
                "given": "Michael"
            }
        ],
        "URL": "https://omanscience.com/ar/articles/capable-capability-aware-policy-adaptation-via-behavioral-latent-encoding",
        "language": "en",
        "issued": {
            "date-parts": [
                [
                    2026
                ]
            ]
        },
        "abstract": "Vision-language-action (VLA) policies assume the embodiment on which they were trained and can fail when a joint fault changes how commanded actions are physically executed. Existing fault-recovery methods often require task-specific retraining, fault labels, explicit diagnosis, or privileged embodiment information. We introduce CAPABLE, a unified capability-aware adaptation framework for frozen VLAs that integrates self-supervised capability inference with residual reinforcement learning. CAPABLE infers capability, how much of the commanded motion each joint actually realizes and how that motion contributes to end-effector behavior, online from command-response history and kinematics using a temporal encoder shared across joints, Jacobian grounding, cross-joint attention, and self-supervised physical prediction. The resulting representation conditions a residual policy that adds bounded corrections to the VLA arm action without fault labels or faulty-joint identifiers. Across 28 LIBERO tasks, CAPABLE raises success on an actuator excluded from fault training from 24.8% to 59.3%, outperforming a parameter-matched global-history baseline by 17.4 points while preserving healthy performance. Leave-one-actuator-out experiments across six joints show that this transfer is not specific to one actuator, and additional evaluations characterize transfer to unseen fault families and demonstrate recovery on a physical Franka Panda. https://capable-vla.github.io/"
    }
]