From Abstraction to Instantiation: Learning Behavioral Representation for Vision-Language-Action Model

Vision-Language-Action (VLA) models often suffer from performance degradation under distribution shifts, as they struggle to learn generalized behavior representations across varying environments. While existing approaches attempt to construct behavior representations through action-centric latent variables, they are often limited by short-horizon temporal fragmentation and static execution-alignment, leading to inconsistent behaviors in complex scenarios. To address these limitations, we propose \textbf{BehaviorVLA}, a framework that facilitates robust manipulation through the learning of a temporally coherent behavioral representations. Our approach features two symmetric components: (1) the \textbf{Visuomotor Behavior Encoder (VBE)}, which utilizes a causal Mamba-based architecture to aggregate long-horizon trajectory information into a unified behavior representation; and (2) the \textbf{Phase-conditioned Behavior Decoder (PBD)}, which decodes this representation into precise actions by dynamically aligning task-level priors with real-time execution progress. Experiments on RoboTwin 2.0, LIBERO, and CALVIN demonstrate state-of-the-art success rates of 58\%, 98\%, and 4.36 (Avg.Len), respectively. Notably, in real-world sim-to-real transfer, BehaviorVLA matches the performance of OpenVLA-OFT using only 50\% of the demonstration data, showcasing its superior data efficiency and generalization.

Robots often struggle when they face new environments, such as different backgrounds, object positions, lighting conditions, or object appearances. This is because many robot learning models tend to memorize visual details from training data rather than learning the underlying way a task should be performed. In this work, we introduce BehaviorVLA, a robot learning framework that helps robots learn reusable behavior patterns from demonstrations. Instead of directly predicting actions only from the current image and instruction, the model first captures the general structure of a task, such as how objects should be moved or arranged, and then adapts this behavior to the robot’s current situation during execution. Experiments in both simulated benchmarks and real-world robot tasks show that BehaviorVLA improves robustness and data efficiency. It performs well under changes in scenes and object layouts, and it can transfer to real-world tasks with fewer demonstrations. These results suggest that teaching robots reusable behavior patterns is a promising way to build more reliable and adaptable robotic systems.