Body Discovery of Embodied AI

TL;DR

This work introduces a new challenge, termed "Body Discovery of Embodied AI", focusing on tasks of recognizing embodiments and summarizing neural signal functionality, and applies causal inference method and evaluates it by developing a simulator tailored for testing algorithms with virtual environments.

Abstract

In the pursuit of realizing artificial general intelligence (AGI), the importance of embodied artificial intelligence (AI) becomes increasingly apparent. Following this trend, research integrating robots with AGI has become prominent. As various kinds of embodiments have been designed, adaptability to diverse embodiments will become important to AGI. We introduce a new challenge, termed "Body Discovery of Embodied AI", focusing on tasks of recognizing embodiments and summarizing neural signal functionality. The challenge encompasses the precise definition of an AI body and the intricate task of identifying embodiments in dynamic environments, where conventional approaches often prove inadequate. To address these challenges, we apply causal inference method and evaluate it by developing a simulator tailored for testing algorithms with virtual environments. Finally, we validate the efficacy of our algorithms through empirical testing, demonstrating their robust performance in various scenarios based on virtual environments.

Figures (6)

• Figure 1: The ability to recognize embodiments is meaningful to embodied AI agents. With such ability, agents can adapt to various sets of embodiments through self-experimentation when being put into new scenarios.
• Figure 2: Framework of the body discovery challenge. The design of the agent is assumed to be invariant across different tasks. Composed of a computational module (i.e., AI mind) and a perception module, the agent is compatible with various embodiments and can be implemented in different world settings. The category of embodiments can be different or even time-varying.
• Figure 3: Simulated scenes for T0-T8. These tasks correspond to the 9 tasks in Table \ref{['tab:scenes']}.
• Figure 4: Results of parametric analysis. Performance changes along seven parameters on T8: Q (neural signal number), N (candidate object number), T (total stage number), N1 (environmental noise intensity), N2 (other agent noise intensity), N3 (action failure intensity), and N4 (sensing flaw intensity). The intensity of N1 and N2 is a calculated ratio. It indicates the range of changes in object features caused by the noise. This ratio is calculated by comparing the changes caused by noise and the changes caused by the neural signals. The intensity of N3 represents the probability of failure each time an action is performed. Thus, it ranges from 0 to 1. The intensity of N4 represents the proportion of sensing error in the actual sensing result each time a perception occurs. Note that in the figure of Q, the lines of precision and specificity almost overlap.
• Figure 5: Simulated scenes for "Mirror Test", i.e., T9-T12. Their settings are similar to T1-T4, and the only difference is that each task has an additional mirror in the scene.

Theorems & Definitions (2)

• Remark 1
• Remark 2