Current frontier humanoid robotic research aims to develop robust control and learning and can be applied in many fields. For example, humanoid robots can be used as autonomous service assistants (healthcare and hospitality), industrial (logistics and manufacturing), and hazardous-setting exploration (wildfire and defense). In these settings, pure humanoid robot learning algorithms would encounter different problems. Firstly, these real-world applications pose great challenges to current humanoid robot learning algorithms, which lack the robustness to perform in long-horizon tasks. Secondly, in uncertain scenarios, robots might misbehave, leading to reduced safety measures that potentially affect humans around them. Lastly, training a humanoid robot policy relies heavily on expert demonstrations, which incur high cost, time, and effort. Furthermore, state-of-the-art robot learning algorithms depend on imitation learning from experts, thus suffering from dataset distribution shift because real-world dynamics differ from those in past demonstration datasets.

To address these challenges, we propose a collection of brain-inspired methodologies that construct the robot’s internal perception and action model. This helps the agent learn the representation of the world in which it interacts with, continuously predicting future outcomes, thus enabling it to optimize its action sequences based on the predicted results of those outcomes. The learning of a perception model is robust to long-horizon tasks because the agent learns to understand the state of the world instead of simply learning to map observations to actions. Additionally, the learning of the world model helps estimate the amount of uncertainty in the agent’s internal model/brain, thus guiding the action sequences to safer policies or exploratory behaviors. Finally, in contrast to the training-then-deploying cycle like current SOTA methods do, our work perception model learns in real-time to understand its own world, enhancing the robustness to dataset distribution shift.

Upon addressing three previously mentioned challenges, humanoid robotics would have more accessibility and robustness in real-world applications. For example, a humanoid robot learning to explore can explore effectively in a search-and-rescue operation thanks to high-uncertainty-based exploration. A humanoid robot while in manufacturing operations, with perception modeling, can constrain its action that it predicts leading to high uncertainty, potentially leading to causing damage.