Birds and experienced glider pilots frequently use atmospheric updrafts for long-distance flight and energy conservation, with harvested energy from updrafts serving as the foundation. Inspired by their common characteristics in autonomous soaring, a reinforcement learning algorithm, the Twin Delayed Deep Deterministic policy gradient, is used to investigate the optimal strategy for an unpowered glider to harvest energy from thermal updrafts. A round updraft model is utilized to characterize updrafts with varied strengths. A high-fidelity six-degree-of-glider model is used in the dynamic modeling of a glider. The results for various flight initial positions and updraft strengths demonstrate the effectiveness of the strategy learned via reinforcement learning. To enhance the updraft perception ability and expand the applicability of the trained glider agent, an extra wind velocity differential correction module is introduced to the algorithm, and a strategy symmetry method is applied. Comparison experiments regarding round updraft, the Gedeon thermal model, and Dryden continuous turbulence indicate the crucial role of the further optimized methods in improving the updraft-sensing ability of the reinforcement learning glider agent. With optimized methods, a glider trained in a simplified thermal updraft with a simple training method can have more effective flight strategies.

Machine learning has become an emerging paradigm for microrobotics, enabling autonomous micro-/nanorobot navigation in complex and highly disturbed environments without requirements of precise models. However, the state-of-the-art learning-based methods adopt “black-box” neural networks (NNs), which raises critical concerns about the trustworthiness of the generated navigation policies, especially for biomedical scenarios. Motivated to address this issue, we propose a physics-constrained learning framework that embeds deterministic physical laws into network architectures to achieve trustworthy microrobot navigation. Instead of learning from scratch, we explicitly encode the monotonic relationships of kinematics and safety rules into NNs. These constraints serve as structural inductive biases, theoretically guaranteeing that navigation policies operate strictly within the trustworthy action domains. Experiments demonstrate that microrobots can autonomously reach random targets without collision with obstacles during long-time and long-distance navigation, proving our framework’s reliability owing to its explainable nature. This work can bridge the gap between data-driven performance and rigorous safety requirements, constituting a meaningful step toward trustworthy artificial intelligence–empowered microrobotics.