Skip to main content
Back to timeline
arXivSource publication:

Long-horizon robot control stops asking the model every step: RoC calls its agent on demand, reaching 99.6% on RMBench with 65.3% fewer tokens

Synopsis

Frequent calls to a multimodal reasoning agent make long-horizon robot manipulation expensive in latency and tokens. RoC uses an online task-state monitor to estimate progress within the active subtask, predict transition requests and propose the next goal, triggering agent review only on progress anomalies, transition requests or critical contact events. On the nine RMBench tasks, RoC averages 99.6% success, 24.0 points above the strongest agent-based baseline with 65.3% fewer tokens and 15.1 points above the strongest policy-based baseline; on the real SO-ARM101 platform its three tasks reach 70%, 80% and 70%.

Source-provided article image: Reasoning on Call: Efficient Agentic Robot Control for Long-Horizon Robot Manipulation
arXiv

Interpretation

Agent calls in long-horizon manipulation move from per-step review to event-triggered review, leaving routine execution to the low-level policy and reasoning to moments that need judgment. Earlier agentic control methods such as GPT-as-Policy repeatedly invoke the agent to review policy proposals, adding reasoning overhead even during routine execution. Across the nine RMBench tasks, RoC averages 99.6% success, 24.0 points above GPT-as-Policy and 27.7 points above direct control, with 65.3% and 84.4% lower token consumption; it makes 46 agent calls versus 84 for GPT-as-Policy and 209 for direct control.

The system has three layers: a low-level policy, an online task-state monitor and a reasoning agent, with the monitor deciding when execution is interrupted. The monitor is initialized from the general-purpose robotic reward model Robometer and jointly estimates within-subtask progress, transition-request probability and a candidate successor goal. The monitor combines stage-aware sampled visual history with ordered interaction memory and is trained for one epoch across all nine RMBench tasks; at inference it runs at 7.10 Hz without subtask generation and 0.53 Hz when generating a subtask, for an overall measured throughput of 2.77 Hz.

Agent review covers execution repair and plan revision together, rather than only deciding whether the subtask should change. Execution repair includes prefix shortening, local adjustment of end-effector poses or gripper commands, and agent takeover of a subtask; plan revision includes rejecting a transition and replacing an invalid successor goal. In the ablation, policy only averages 5.9% success, adding visual history raises it to 14.7%, adding the monitor to 63.7%, and full RoC reaches 98.1%; on memory-dependent tasks the monitor-equipped policy averages 51.0%, rising to 97.3% with event-triggered agent intervention.

Perspective

The demonstrated scope is all nine simulated RMBench tasks and three tabletop tasks on the SO-ARM101 platform, Drawer Sorting, Cover and Reveal and Tabletop Cleanup, with 10 trials per real task; it suits robotics researchers working on long-horizon manipulation and agent scheduling.

Real-robot failures mainly involve difficult grasps that remain unsuccessful after agent intervention, where the current reasoning agent fails to turn visual evidence into effective grasp corrections, leaving agent capability as a bottleneck; the authors suggest a stronger agent could be integrated without redesigning the control framework but report no experiment with such a replacement.