This work presents a hierarchical reinforcement learning (RL) framework that empowers quadruped robots to execute object manipulation tasks through body-assisted interactions, without relying on traditional end-effectors. The proposed approach replaces traditional reliance on grippers or privileged, simulation-specific data with a sensor-driven control architecture designed for real-world applicability. The framework employs a streamlined design using fully connected deep neural networks to model both high-level and low-level controllers, eliminating the need for computationally intensive components such as LSTM encoders and privileged latent inputs. This simplification reduces computational overhead while maintaining robust functionality in diverse operational environments. The system’s performance was evaluated through simulations in NVIDIA Isaac Sim, using quadruped robots developed by ANYbotics. The experimental findings indicate that the proposed method achieves high-precision object manipulation, with an average final positioning error of 13 cm across object-to-target distances extending up to 10 m. The employed reinforcement learning framework effectively reduces control effort, thereby promoting energy-efficient operation and enhancing stability throughout task execution. The findings demonstrate the framework’s potential as a foundational methodology for advanced robotic applications requiring adaptability and Practical applicability.