The article introduces InsActor, a generative framework for creating physics-based animations driven by high-level human instructions. The framework leverages recent advancements in diffusion-based human motion models to capture complex relationships between instructions and character movements. InsActor addresses issues of invalid states and infeasible transitions in planned motions by discovering low-level skills and mapping plans to latent skill sequences. The framework has demonstrated its effectiveness in tasks such as instruction-driven motion generation and waypoint heading. The use of high-level instructions makes InsActor particularly useful for executing long-horizon tasks with a rich set of instructions.
Publication date: 28 Dec 2023
Project Page: jiawei-ren.github.io/projects/insactor
Paper: https://arxiv.org/pdf/2312.17135
