Move-Then-Operate: Behavioral Phasing for Human-Like Robotic Manipulation

Abstract: We present Move-Then-Operate, a Vision language action framework that explicitly decouples robotic manipulation into two distinct behavioral phases: coarse relocation (move) and contact-critical interaction (operate). Unlike monolithic policies that conflate these heterogeneous regimes, our architecture employs a dual-expert policy routed by a learnable phase selector, introducing a structural inductive bias that isolates phase-specific dynamics. Phase labels are automatically generated via an MLLM-based pipeline conditioned on lightweight contextual cues such as end-effector velocity and subtask decomposition to ensure alignment with human motor patterns. Evaluated on the RoboTwin2 benchmark, our method achieves an average success rate of $68.9\%$, outperforming the monolithic $π_0$ baseline by $24\%$. It matches or exceeds models trained on $10\times$ more data and reaches peak performance in $40\%$ fewer training steps, demonstrating that architectural disentanglement of move and operate phases is a highly effective and efficient strategy for mastering high-precision manipulation.

• The paper presents a dual-expert policy that explicitly decomposes robotic manipulation into move and operate phases.
• It employs Conditional Flow Matching with a phase-aware annotation pipeline to improve training efficiency and precision.
• Empirical results on the RoboTwin2 benchmark show significant success rate improvements and reduced training steps.

Behavioral Phase Disentanglement for Robotic Manipulation in Move-Then-Operate

Motivation and Context

Robotic manipulation tasks inherently involve two qualitatively distinct behavioral regimes: rapid, coarse relocation to bring the end-effector close to a target ("move"), and precise, contact-critical actuation to achieve task-specific interaction ("operate"). Human motor control, characterized by this two-phase process, motivates a structural decomposition that is largely unexploited in prevailing monolithic Vision-Language-Action (VLA) models. These models either optimize both regimes under a unified objective or apply coarse hierarchical decomposition that does not prevent gradient interference across phases. This failure to separate move and operate phases prevents exploitation of phase-specific dynamics and leads to suboptimal data efficiency, particularly for high-precision manipulation.

Architectural Contributions

Move-Then-Operate proposes an explicit phase decomposition for manipulation via a dual-expert policy gated by a learnable phase selector. Phase selection is performed at the chunk (control step) level, mapping the multi-modal context—encoded by a shared vision-language backbone—into phase logits using masked global average pooling to generate semantic scene representations, enabling robust and temporally consistent routing. Each expert instantiates a Conditional Flow Matching (CFM) model, operating over disjoint parameter sets for move and operate phases. This explicit separation counters the optimization instability where the prevalence of large-amplitude move actions masks the learning signal for the operate regime under joint normalization.

The routing strategy employs ground-truth phase labels for supervised teacher-forcing during training, and during inference, expert selection follows a greedy decoding over router predictions. This hard routing enforces orthogonalization of update signals, yielding specialized vector fields for each control regime. The integration of CFM provides smooth interpolation between initial latent state and target action, parameterized for the specific phase context.

Phase-Aware Annotation Pipeline

A phase-aware annotation pipeline underpins the training efficiency of the approach. This system leverages a Multimodal LLM (MLLM) to segment demonstration trajectories into temporally contiguous move and operate phases, informed by contextual cues (task decomposition, velocity heuristics) and explicit topological constraints. Annotation proceeds via iterative MLLM prompting and deterministic validation, enforcing structural validity while ensuring phase assignments align with human demonstration statistics. The iterative correction loop refines labels via structured error feedback, raising phase label quality above standard heuristic or unsupervised methods.

Empirical Performance

Evaluation on the RoboTwin2 benchmark (eight representative contact-rich and transport tasks, with 50 demonstrations per task and 100K training steps) demonstrates:

• A +24.1% absolute success rate improvement (68.9%) over a strong monolithic To baseline.
• Data efficiency matching or exceeding “data-rich” baselines trained with 10× more demonstrations: For instance, achieving 99% success on “Click Bell” and 93% on “Press Stapler”, outperforming models with far larger training corpora, especially on contact-intensive tasks.
• Training efficiency gains: Peak performance is reached in 40% fewer training steps compared to monolithic baselines. Rapid convergence is particularly evident in fine-tuning regimes for high-precision or long-horizon manipulation.
• Ablation reveals critical dependency on accurate phase routing: Random or adversarial routing leads to drastic performance drops (down to <9% average success, versus 68.9% for “original selection”), demonstrating lack of cross-expert generalization and strongly phase-specialized policy learning.

Success rates for high-precision tasks improve up to +55% over baselines, substantiating the hypothesis that conflation of move/operate phases constrains policy optimality under standard joint training.

Implications and Theoretical Considerations

This separation of behavioral phases introduces a structural inductive bias that directly addresses the heterogeneity of manipulation dynamics, providing several theoretical and practical implications:

• Optimizing Disjoint Vector Fields: By phase-conditional learning, the experts avoid conflicting gradient signals and can capture statistics otherwise suppressed or distorted in joint modeling.
• Efficient Credit Assignment in Imitation Regime: The architecture enhances visibility for rare or low-amplitude contact events, mitigating the “drowning out” of fine-grained behaviors by more frequently occurring move actions.
• Scaling Laws for Data Use: Results suggest that architecture choice can outweigh brute-force scaling of data quantity for mastery of complex, underrepresented behaviors—an important insight for generalist robotics under real-world data constraints.
• Automated Labeling and Generalization: The automated, context-aware labeling pipeline provides a scalable methodology for phase-conditioned learning that is likely to generalize across task families and embodiments.

Future Directions

The Move-Then-Operate framework sets a precedent for further structural decompositions in VLA models. Potential research directions include: dynamic gating at finer temporal scales, joint optimization with RL signals in multi-expert settings, phase-aware curriculum learning, and extension to multi-phase or multi-modal division reflecting additional task complexity (e.g., force modulation, multi-contact manipulation). Integration with continual or open-world learning platforms could further elucidate the long-term generalization benefits of phase-disentangled controllers.

Conclusion

Move-Then-Operate demonstrates that explicit behavioral phasing, realized as phase-conditional dual-expert flow matching with automated annotation, provides substantial and robust gains over monolithic policies for high-precision robotic manipulation. Architectural disentanglement, more than increased parameter count or data scale, emerges as a decisive strategy for mastering complex manipulation under realistic data and training constraints. These findings motivate continued exploration of structural biases in policy learning for embodied agents and highlight the importance of task-aware decomposition in practical robotics.