Light-WAM: Efficient World Action Models with State-Fusion Action Decoding
Researchers have developed Light-WAM, a more efficient model for robot policy learning that incorporates future prediction. This new model uses a compact video backbone and performs future-video supervision in a downsampled latent space, significantly reducing training costs. Light-WAM also features a StateFusionActionExpert that fuses adapted states from multiple backbone layers to directly predict action chunks, leading to faster inference and lower memory usage while maintaining strong performance on manipulation tasks. AI
IMPACT This model offers a more efficient approach to robot policy learning, potentially enabling wider deployment of advanced robotic systems.