Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control
Agile-WAM reports a 29.4% relative success-rate gain in real-world contact-rich manipulation while running at 11.9 ms latency.
The paper introduces a tactile World Action Model that predicts actions and future visual/tactile states without relying on large pretrained generative backbones. Its key design separates timing: visual latents are supervised farther ahead, while tactile latents are predicted at the next frame to catch abrupt contact changes. The authors test it on nine simulated and five real-world manipulation tasks and say it beats the strongest baseline in success rate. ArXiv · AI/CL/LG's note
The paper introduces a tactile World Action Model that predicts actions and future visual/tactile states without relying on large pretrained generative backbones. Its key design separates timing: visual latents are supervised farther ahead, while tactile latents are predicted at the next frame to catch abrupt contact changes. The authors test it on nine simulated and five real-world manipulation tasks and say it beats the strongest baseline in success rate. ArXiv · AI/CL/LG's note
score 5