Flex-$π$: A Multi-Stream World-Action Model with Compute Flexibility

FLEX-π is a 6B-parameter world-action model that jointly predicts robot actions with future RGB, 3D pointmaps, and object-centric DINOv3 features. Its key finding is that the frozen Wan-2.2 VAE, trained for RGB video, reconstructs pointmaps accurately enough to provide geometric supervision without new sensors or VAE pre-training, while Depth Anything 3 supplies pointmaps from RGB and DINOv3 supplies semantic tokens. A Mixture-of-Transformers backbone fuses these streams, and independent 0.5-probability visual dropout with cross-modality forcing trains one checkpoint to operate from any subset of inputs and generate any subset of futures, from action-only to full joint inference. Pre-trained on 500 hours of AGIBOT World, FLEX-π reaches 94.6% randomized success on the 50-task RoboTwin benchmark, and its fixed full-joint variant reaches 99.2% on LIBERO, matching Qwen-RobotManip while using far less robot data. On a real bimanual YAM robot, it outperforms π0.5, ManiFlow, and Fast-WAM on five dexterous tasks; action-only inference runs at 60 ms per call and already exceeds the baselines, while joint visual generation raises average success by 13%. The additional geometry stream is consequential: removing pointmaps reduces RoboTwin success by 20%, and cross-modality forcing improves representation learning even when modalities are available at deployment.