Large multimodal models can support long-horizon robot manipulation, but frequent model calls incur substantial latency and token costs. The central challenge is determining when agent assistance or intervention is needed and whether errors require correcting robot execution, subtask planning, or both. To address this problem, we propose RoC (Reasoning on Call), an efficient framework that coordinates a low-level policy and a high-level reasoning agent through an online task-state monitor. The monitor combines temporal visual observations with ordered interaction memory to jointly estimate progress within the active subtask and predict its successor. These predictions guide event-triggered agent review for coordinated execution repair and plan revision. Verified outcomes govern memory updates and subtask transitions, allowing routine execution to remain with the policy while reasoning focuses on decisions requiring additional assessment. On RMBench, RoC improves task success rate by 24.0 percentage points and reduces token consumption by 65.3% compared with the strongest agent-based baseline. It also outperforms the strongest policy-based baseline by 15.1 percentage points in task success rate. Real-world experiments further validate its effectiveness across three long-horizon manipulation tasks. Overall, RoC supports efficient agent intervention and coordinated recovery across execution and planning for long-horizon robot manipulation.
Yuanzhi He, Victor Romero-Cano, Jiayu Xu, Juan David Hernández, Gualtiero Colombo
原文摘要
Obstacle-aware robotic manipulation can require substantial interaction data when deep reinforcement learning policies are trained from scratch. We study a task decomposition framework combining first-layer critic transfer with staged reward redesign (RR) on a Truncated Quantile Critics (TQC) backbone. Transfer reuses an input representation across tasks, while RR supplies intermediate feedback for Pick-and-Place. Experiments cover Push, Pick-and-Place, and Slide in obstacle-free and single-obstacle PyBullet environments, with three random seeds per setting. On obstacle-aware Push, transfer achieves $0.83\pm0.05$ success against $0.77\pm0.06$ for TQC. On obstacle-aware Slide, it achieves $0.95\pm0.02$ against $0.84\pm0.06$ for SAC, the strongest baseline in that setting. These are gains of $6$ and $11$ percentage points, respectively. RR reduces the target-training steps needed for obstacle-free Pick-and-Place from roughly $0.75$M to $0.20$M. Transfer benefits depend on the route: obstacle-aware Pick-and-Place shows a lower mean than training from scratch. A selected simulated trajectory is executed open-loop on a Franka Emika FR3 as a feasibility demonstration. The study characterizes when a lightweight combination of established transfer and reward-design components helps learning.