W2-VLA forecasts wrist motion for finer robot control

W2-VLA, introduced by Yuhao Pan and colleagues, tackles a core flaw in vision-language-action (VLA) models: they usually treat main-view and wrist-view images as equal inputs. In fine-grained manipulation, the wrist camera's local perspective matters differently—it should be predicted based on the global task. W2-VLA builds a compact interface of latent modeling tokens between the vision-language model and a wrist predictor. Given the current multi-view observations and a task instruction, the predictor uses these tokens plus observed wrist history to forecast future wrist latents, which become future-aware context for action prediction. The companion W2-CoT pipeline generates structured annotations describing manipulation progress, physical transition cues, and wrist-local evidence, providing auxiliary supervision that shapes the latent interface.

On benchmarks and hardware, W2-VLA shows clear gains on LIBERO, RoboTwin 2.0, and real-world manipulation setups. It excels in fine-grained, contact-sensitive tasks—both single-arm and bimanual—translating to more precise grasping, assembly, and tool use. Critically, these improvements come without sacrificing speed: the model sustains action-generation rates above 80 Hz, fast enough for closed-loop robot control. By separating the roles of world view and wrist view, W2-VLA gives robots a better hand-eye coordination model, making it a promising step toward truly dexterous manipulation in unstructured environments.