Integrating Agglomerative Perception with One-step Action Generation for Robotic Manipulation.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Developing generalizable robotic policies faces challenges in balancing efficiency, accuracy, and robustness.
• Existing Vision-Language-Action models require extensive data, while keyframe methods struggle with generative model expressivity and sampling latency.

Purpose of the Study:

• To present Flow2Act, a unified framework addressing the trilemma of robotic policy development.
• To enable efficient, accurate, and robust robotic manipulation through a novel approach.

Main Methods:

• Introduced an agglomerative multi-teacher visual backbone for robust representations without task-specific pretraining.
• Proposed a conditional MeanFlow policy for deterministic single-step action generation, avoiding ODE-based flow matching issues.
• Devised a curriculum region-aware mechanism using Spatial-Grounded State Space Duality for progressive attention shift.

Main Results:

• Flow2Act demonstrated significant gains in policy performance on simulation benchmarks and real-world robotic tasks.
• The framework showed improved robustness to environmental perturbations.
• Validated cross-task real-world applicability of the developed robotic policies.

Conclusions:

• Flow2Act offers a unified framework overcoming limitations of existing methods for robotic policy development.
• The approach achieves a balance between inference efficiency, manipulation accuracy, and robustness.
• Significant improvements in performance and applicability were observed in real-world robotic tasks.