PROJECT & FUNDING SCHEMES

General purpose humanoids are expected to play a key role in future industries, homes, healthcare, and logistics, serving as reliable assistants that enhance safety, efficiency, and human wellbeing. Forecasts suggest that the market scale in China will exceed 60 million units by2050. However, current models are still restricted to predefined motions such as walking, running, waving, or jumping and lack robust generalpurpose manipulation skills. Key challenges include unstable locomotion when carrying heavy objects, which can cause excessive stopping distances and safety risks, as well as limited autonomous perception that leads to dependence on manual control. To address these limitations, we propose a dual agent reinforcement learning framework trained in Isaac Sim. The first agent focuses on force adaptive locomotion, supported by a progressive torque and force aware curriculum that improves balance when handling heavy loads. The second agent enables intelligent object manipulation through the Segment Anything Model, six dimensional pose estimation, and continuous tracking, allowing autonomous object recognition, grasp point selection, and execution through inverse kinematics.
This project seeks to advance humanoids from preprogrammed machines to adaptive partners. The benefits include improved stability, higher safety, and autonomous decision making, enabling real world deployment in logistics and manufacturing where human like adaptability and efficiency are essential.