EXACT Language Models Human Motion as Executable Programs for Zero-Shot Policy Inference
A recent research paper presents EXACT, a specialized language aimed at depicting human movements through underspecified motion programs. These programs serve as reward-generating functions for zero-shot policy inference, utilizing forward-backward representations. This method seeks to elucidate the internal dynamics and framework of human actions, essential for evaluating execution quality and distinguishing between different actions. Current action recognition techniques often overlook the execution process, concentrating solely on the actions themselves. Human-centered systems necessitate a grasp of human actions in the physical realm, where intentional and structured sequences unfold over time. By exploiting the compositional characteristics of EXACT motion programs, distinct policies merge into an executable neuro-symbolic model that employs program structure for compositional analysis. The effectiveness of this proposed framework for comprehending human actions is assessed in the paper, which was published on arXiv under identifier 2604.18064v1.
Key facts
- EXACT is a domain-specific language representing human motions as motion programs
- Programs are interpreted as reward-generating functions for zero-shot policy inference
- Forward-backwards representations are used in the inference process
- The approach captures internal mechanics and structure of human actions
- Existing action recognition methods often miss execution structure
- Human-centered systems require understanding of physical world actions
- Motion programs combine into executable neuro-symbolic models
- Research announced on arXiv with identifier 2604.18064v1
Entities
Institutions
- arXiv