In the face of current modular robots that rely on manual coding to generate deformation instructions and are limited by binary instructions, this research proposes a Cognitive Modular Control (CMC) architecture inspired by ACT-R theory. Build simulation tool ReoForm to conduct experiments through simulation-based experimental design. The data results show that each module of the architecture demonstrates feasibility and universality in terms of time-to-completion, energy consumption, manipulability index and robustness indicators.

To improve the cognitive autonomy of humanoid robots, this research proposes a multi-scenario reasoning architecture to solve the technical shortcomings of multi-modal understanding in this field. It draws on simulation based experimental design that adopts multi-modal synthesis (visual, auditory, tactile) and builds a simulator "Mahā" to perform the experiment. The findings demonstrate the feasibility of this architecture in multimodal data.

To address the technical bottleneck caused by the domain mismatch problem in the sim-to-real rransfer process of humanoid robots, a targeted domain refinement (TDR) framework is proposed to solve the shortcomings of adaptive domain randomization (ADR). This research design uses quantitative simulation experiments to execute the code of TDR and ADR architecture through the Python 3.13 IDLE platform. Then the sample efficiency, adaptation speed, robustness, policy stability and domain gap of synthetic data are used to compare and analyze the two frameworks. The results show that TDR outperforms ADR on five indicators. It proves that it can make up for the shortcomings of ADR in solving domain mismatch problems, and provides reference experience for future research on improving sim-to-real rransfer.