Editorial: Reinforcement learning for real-world robot navigation

Path planning is essential for deploying autonomous robots in complex real-world environments, yet long-term decision-making and sparse rewards pose significant challenges for traditional reinforcement learning (RL) algorithms. Offline hierarchical reinforcement learning (HRL) addresses these issues by decomposing tasks into highlevel subgoal generation and low-level subgoal attainment. Advanced methods like Guider and HIQL introduce latent spaces to represent subgoals, but their high-level policies must search over continuous latent spaces-a process that remains sampleinefficient and often leads to unstable training, especially with policy gradient algorithms like PPO. To overcome this limitation, Xiang Han proposes LG-H-PPO, a novel offline hierarchical PPO framework that discretizes the continuous latent space into a structured latent graph. This transforms high-level planning from difficult "continuous creation" into simple "discrete selection," substantially reducing learning complexity. Preliminary experiments on D4RL benchmarks show that LG-H-PPO outperforms strong baselines in convergence speed and success rate. The key contribution is integrating graph structures into latent-variable HRL planning, which simplifies the high-level action space and improves training efficiency and stability for long-sequence navigation tasks.Developing reliable deep reinforcement learning (DRL) navigation policies for mobile robots in highly dynamic, real-world environments remains challenging, especially when robots must explore unknown spaces without prior knowledge while avoiding collisions. Bockrath et al. introduces Trust-Nav, a trustworthy navigation framework that uses variational policy learning to quantify uncertainty in action estimation, localization, and map representation. By applying Bayesian variational approximation to the policy network's parameters and combining policy-based with value-based learning, Trust-Nav propagates variational moments through all network layers. Uncertainty in robot actions is measured via propagated variational covariance, while uncertainty in localization and mapping is embedded into the reward function using optimal experimental design theory. Experiments in the Gazebo simulator demonstrate that Trust-Nav achieves robust autonomous navigation and mapping, consistently outperforming deterministic DRL approaches, particularly under noisy conditions and adversarial attacks. By integrating uncertainty into the policy network, Trust-Nav promotes safer, more reliable navigation and represents a step toward selfaware robotic systems that can recognize and respond to their own limitations. In multi-skill imitation learning for robots, complete expert datasets with full motion features are essential but often unavailable. Datasets based solely on joint positions are more accessible yet lack the detail needed for effective skill learning and smooth transitions. To address this, Tu et al. proposes Seamless Multi-Skill Learning (SMSL) framework. Built upon the Adversarial Motion Priors framework and enhanced with self-trajectory augmentation, SMSL leverages high-quality historical experiences to guide skill acquisition and generate natural transitions, overcoming the limitations of incomplete data. An adaptive command sampling mechanism balances training across skills of varying difficulty and prevents catastrophic forgetting. Experiments show that SMSL outperforms baseline methods, and sim-to-real validation on Solo8 robots confirms its effectiveness. This work demonstrates SMSL's potential for autonomous skill learning from minimal data in real-world robotic applications. Traditional wilderness search and rescue methods are slow and limited in coverage. Drones offer speed and flexibility, but effective operation requires optimized search paths. Ewers et al. presents a deep reinforcement learning algorithm that generates efficient drone search paths using a probability distribution map of the search area and missing person. The learned policy maximizes the probability of rapid detection. Experimental results show that the proposed method reduces search times by over 160% compared to traditional coverage and search planning algorithms-a critical difference in real-world rescue operations. Furthermore, unlike prior work, this approach employs a continuous action space enabled by cubature, allowing more nuanced and adaptive flight patterns. Soft robots offer compliance and human-safe interaction in unstructured environments, but their soft structures make control algorithm development challenging. Oikonomou et al. introduces a novel motion control technique for a modular bio-inspired softrobotic arm, focusing on qualitative reproduction of periodic trajectories. The method combines Probabilistic Movement Primitives (ProMP) and Central Pattern Generators (CPG): ProMP provides immediate actuation estimation using a learned library of simple movements without time-consuming training, while CPG filters and parameterizes the resulting signals to generate rhythmic patterns at the motor level. Evaluated on a two-module soft arm, the approach enables rapid acquisition of periodic movements and compression into low-dimensional CPG parameters for longterm storage and execution.In summary, the works reviewed highlight significant advances in applying reinforcement learning and related paradigms to robot navigation and control. Key themes include the integration of hierarchical learning to handle long-term sparse rewards, the use of uncertainty quantification for trustworthy decision-making, the combination of movement primitives with rhythmic pattern generators for soft robot control, and the fusion of deep RL with prior knowledge such as probability maps for drone search to dramatically improve efficiency. Offline hierarchical RL and multiskill imitation learning further address data scarcity and incomplete demonstrations, while foundation models offer new prospects for generalization and multimodal reasoning. Looking ahead, several directions merit attention. First, bridging the gap between simulation and real-world deployment remains critical, as evidenced by sim-to-real validations. Second, combining explicit graph-based planning with latent variable representations could yield more sample-efficient and stable hierarchical policies. Third, embedding uncertainty awareness into all levels of the control stack from perception to action will be essential for safe operation in unpredictable environments. Finally, leveraging foundation models as prior knowledge for low-data adaptation and cross-task transfer presents a promising frontier. We envision that the convergence of these approaches will lead to more autonomous, robust, and deployable robotic systems capable of navigating complex real-world challenges.