From Error Accumulation to Error Contraction: Stable Long-Horizon Embodied Manipulation via Primitive-Space Correction

Abstract

Failures in long-horizon, multi-stage embodied manipulation are often not caused by isolated decision errors, but by the progressive amplification of execution deviations across sequential stages, which ultimately destabilizes the overall process. To address this issue, we investigate the problem from the perspective of error propagation dynamics and propose a predictive–residual-triggered execution framework operating in the primitive parameter space. The proposed approach performs forward evaluation within a low-dimensional structured action manifold and activates local parameter re-optimization only when the residual between predicted and observed states exceeds a predefined threshold. By constraining online corrections to the primitive parameter space, the framework suppresses error accumulation without relying on high-frequency global replanning. We evaluate the method through extensive real-world robotic experiments on three representative manipulation tasks, including pick-and-place, tool use, and long-horizon precision assembly, together with targeted simulations for controlled disturbance analysis. The real-world results demonstrate more stable stage-wise success profiles and bounded residual growth, while the targeted disturbance analysis further indicates improved robustness under controlled perturbations. These results suggest that regulating error propagation at an appropriate action representation level is critical for achieving stable long-horizon embodied execution.