Feature-space residual mixture-of-experts action adapter for long-horizon robot manipulation

Abstract

Long-horizon robot manipulation requires a policy to coordinate heterogeneous action primitives across sequentially issued language instructions, while remaining responsive to evolving visual observations and contact states. Existing Vision-Language-Action (VLA) models have substantially advanced general robotic control. However, when handling multiple semantic goals in sequence, a single action head may fail to capture the distributional diversity across distinct manipulation phases. To this end, we propose the Feature-Space Residual Mixtureof-Experts Action Adapter, which combines the Feature-Space Residual Mixture-of-Experts Action Head with a Prompt-Guided Expert Router at the action-feature interface of a frozen VLA backbone. The action head learns conditional feature-space corrections tailored to distinct manipulation phases through temporal experts. The router, in turn, performs phase-aware expert assignment and employs a prompt-prior confidence coefficient to adaptively modulate the influence of language semantics on routing decisions. Together, these modules enable phase-dependent action specialization without retraining the full backbone. Experimental results on long horizon manipulation benchmarks demonstrate consistent improvements in average successful sequence length and multi-step sequential success rates.