Summary
This article summarizes the titled paper on SImulated testbed for human level AI This paper proposes a 3D Simulated environment for developing generalized intelligent agent. They suggest that the main reason behind non generalized learning agents are-
- Targeting a single task rather than diverse task.
- Use of refined and focused datasets rather than diverse and noisy datasets
- Relying on explicit rewards rather than on other mechanisms
- Too many necessary components rather than a sufficient set of the learning mechanism.
SEDRo provides 2 different environments.
- Fetus: Simulates the womb
- After birth: Simulates 0 to 12 months after birth. The learning agent is a human baby. It also provides a pre-programmed (To interact with the baby and provide the essential experience for human level AI) Social partner.
The proposed environment has 3 unique features:
- Open-ended tasks without extrinsic reward: Provides no reward, no specific goal.
- Human-Like experience with social interaction: Learning agent is given a human baby’s environment. Agent model performance can be compared against human performance.
- Longitudinal development: Unfolds agent capabilities(Motor strength, Vision, etc) according to a curriculum following milestones from developmental psychology.
And finally SEDRo also provides a set of evaluation methods from developmental psychology research.
Concluding thoughts: Overall the paper presents a novel idea. However how the essential experience will be chosen and developed remains a challenge.
