<p>Soft hydrogel microrobots offer unique advantages for safe, minimally invasive biomedical interventions, yet precise control in tissue-like media remains challenging due to weak ultrasound echoes and actuation uncertainty. Here, we use a millimeter-scale magnetic hydrogel robot as a proof-of-concept platform to develop an ultrasound-guided closed-loop control strategy for adaptive navigation of soft hydrogel microrobots. Real-time localization is achieved by combining inter-frame differencing with a correlated-noise Kalman filter, providing stable feedback even under low acoustic contrast. A hierarchical controller—model predictive control (MPC) for swimming and proportional–integral–derivative (PID) control for rolling—executes trajectories and switches gaits under a rigorously validated stability rule. Inspired by the crown-of-thorns starfish, the hydrogel robot adaptively transitions between swimming and rolling in response to local environmental conditions. In heterogeneous tissue-mimicking phantoms, the system achieves sub-4 mm trajectory-tracking error, climbs graded surfaces, clears vertical steps through adaptive gait switching, and explores unseen obstacles without precomputed maps. By integrating ultrasound sensing, adaptive feedback, and multimodal locomotion, this work transforms compliant hydrogels from passive movers into predictable robotic agents, providing a generalizable framework for imaging-integrated control of soft microrobots.</p>