Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
SmolWAM is a latent world-action model designed for efficient robotic manipulation on edge devices, addressing the computational cost of existing video diffusion transformer-based WAMs. By leveraging a pretrained 2B-parameter policy and optimizing its inference interface with single-frame latents, cached text features, a reduced transformer block count, and a fixed token set, SmolWAM achieves significantly faster inference and lower resource utilization. The model demonstrates strong performance, succeeding in 95.5% of LIBERO episodes at 57 ms per chunk while retaining the ability to forecast scene motion.
A 2B-parameter WAM can be distilled to run on edge devices at 11x the speed and 18.5x fewer FLOPs, enabling real-time robotic manipulation with minimal performance degradation.
World-action models (WAMs) predict a robot's actions together with the observations those actions will produce. However, the strongest models rely on video diffusion transformers that are far too computationally expensive for resource-constrained edge devices. We present SmolWAM, a latent WAM that we derive from the 2B-parameter Cosmos Policy by keeping its pretrained trunk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x faster than its teacher with 18.5x fewer FLOPs and 3.2x less memory, and its world head still forecasts scene motion.