Recent advances in quadrupedal robots have demonstrated impressive agility and the ability to traverse diverse terrains. However, hardware issues, such as motor overheating or joint locking, may occur during long-distance walking or traversing through rough terrains leading to locomotion failures. Although several studies have proposed fault-tolerant control methods for quadrupedal robots, there are still challenges in traversing unstructured terrains. In this paper, we propose DreamFLEX, a robust fault-tolerant locomotion controller that enables a quadrupedal robot to traverse complex environments even under joint failure conditions. DreamFLEX integrates an explicit failure estimation and modulation network that jointly estimates the robot's joint fault vector and utilizes this information to adapt the locomotion pattern to faulty conditions in real-time, enabling quadrupedal robots to maintain stability and performance in rough terrains. Experimental results demonstrate that DreamFLEX outperforms existing methods in both simulation and real-world scenarios, effectively managing hardware failures while maintaining robust locomotion performance.

Quadrupedal robots resemble the physical ability of legged animals to walk through unstructured terrains. However, designing a controller for quadrupedal robots poses a significant challenge due to their functional complexity and requires adaptation to various terrains. Recently, deep reinforcement learning, inspired by how legged animals learn to walk from their experiences, has been utilized to synthesize natural quadrupedal locomotion. However, state-of-the-art methods strongly depend on a complex and reliable sensing framework. Furthermore, prior works that rely only on proprioception have shown a limited demonstration for overcoming challenging terrains, especially for a long distance. This work proposes a novel quadrupedal locomotion learning framework that allows quadrupedal robots to walk through challenging terrains, even with limited sensing modalities. The proposed framework was validated in real-world outdoor environments with varying conditions within a single run for a long distance.

Among the challenges in the recent research of end-to-end (E2E) driving, interpretability and distribution shift in the simulation-to-real (Sim2Real) have drawn considerable attention. Because of low interpretability, we cannot clearly explain the causal relationship between the input image and the control actions by the network. Moreover, the distribution shift problem in Sim2Real degrades the driving performance of the policy in the realworld deployment. In this paper, we propose a segmentation-based classwise disentangled latent encoding algorithm to cope with the two challenges. In the proposed algorithm, multi-class segmentation transfers RGB images in both simulation and real environments to the same domain, while preserving the necessary information of objects of primary classes, such as pedestrian, road, and cars, for driving decisions. Besides, in the class-wise disentangled latent encoding, segmented images are encoded to a latent vector, which improves the interpretability significantly, since the state input has a structured format. The interpretability improvement is testified by the t-stochastic neighbor embedding, image reconstruction and the causal relationship between the real images and the control actions. We deploy the driving policy trained in the simulation directly to an autonomous vehicle platform and show, to the best of our knowledge, the first demonstration of the RL-based E2E autonomous in various real environments.