Unitree has unveiled a new humanoid robot, dubbed "Superman," claiming a top running speed of 12.66 m/s and a standing high jump of 2 meters. The robot surpasses the peak human velocity of 12.42 m/s set by Usain Bolt, fulfilling a bold prediction made by Unitree's CEO earlier this year. The prototype has been in development for just over three months and has only just started improving, with the company stating there is "significant room for further improvement in the coming months."
A new AI agent called Roboto Agents has been released, which can analyze robotics data to help find the cause of a failure. The agents can perform tasks such as triage, root cause analysis, data curation, and data analysis, and can work with various types of data including ROS bags, video, and source code. With Roboto Agents, teams can improve their robot runs by analyzing data and identifying areas for improvement. The agents are now available for use and feedback is welcomed.
Researchers Mariana del Castillo and Federico Larroca have proposed an alternative approach to controlling mobile communication agents in a network, called A-Fiedler, which replaces the traditional Fiedler-gradient controller with the dominant eigenvector of the adjacency matrix. This approach provides improved robustness under distributed estimation and maintains network performance even under communication constraints. Compared to the classical Fiedler-gradient controller, A-Fiedler shows comparable network performance in the absence of communication constraints and improved robustness. The new approach offers a simpler path toward distributed network control.
hint$^2$ introduces a method for guiding short-horizon policies toward satisfying complex Linear Temporal Logic (LTL) specifications at inference time using hierarchical world models. Two separate guidance objectives are derived: a high-level model predicts future action-induced transitions in task-relevant atomic propositions, and a low-level dynamics model predicts immediate state evolution for accurate local safety guidance. The method outperforms existing inference-time steering methods in CALVIN and successfully completes instructions with complex liveness and safety constraints more elegantly than language-conditioned alternatives.
Graph-MambaNav: a target-aware spatial-temporal graph encoding framework leverages object-relation knowledge for object-goal navigation, introducing a heuristic ordering over objects based on their relevance to the target, and combining local message passing with global GraphMamba-based selective scanning and sequence modeling over object-wise temporal orders. Experiments on AI2-THOR and RoboTHOR demonstrate improved navigation performance with generalization, and real-world robot deployment further validates the approach's effectiveness. The framework initializes node ordering and edge weights from LLM-derived commonsense object relationships, providing a unified prior for structured reasoning.
Knowledge-Data-Dual-Driven Reinforcement Learning (KDDRL) is proposed to address challenges in autonomous vehicle control in mixed traffic. The proposed method synthesizes intention-aware future trajectories using a conditional deep generative model, fuses probabilistic data-driven insights with physical constraints to guide safe exploration, and compresses both intention-aware trajectories and physical constraints into compact shared embeddings. Evaluations on dataset-calibrated simulations demonstrate that KDDRL effectively handles intention uncertainty, accelerates training convergence, and outperforms conventional baseline methods in terms of safety, efficiency, and comfort.
Researchers developed a modular robotic motor control system inspired by the bilateral hemispheric organization of the brain. The system, which consists of two GRU-based modules connected by a learnable, delayed inter-hemispheric channel, outperforms a monolithic baseline in reaching and holding tasks. The modular architecture improves endpoint precision, reduces muscle co-contraction, and lowers energetic cost, particularly in non-zero-delay regimes.
AdvDex presents a unified Vision-Language-Action framework for learning dexterous manipulation from human and robot demonstrations. It uses a large-scale multimodal dataset, OmniShare, to provide high-quality kinematic supervision and tactile measurements. The framework introduces a canonical action representation, Joint-Aligned Action Space (JAAS), which aligns human hands, dexterous robot hands, and parallel grippers. AdvDex achieves consistent improvements over baselines, effective zero-shot human-to-robot skill transfer, and data-efficient few-shot adaptation.
Researchers demonstrated a space robot teleoperation system over a lossy and delayed network using the Autonomy Testbed for Multi-purpose Orbiting Systems (ATMOS). ATMOS is a planar spacecraft-analog robot designed for testing guidance and control strategies in microgravity-like conditions. The demonstration includes a long-distance remote control experiment between Seoul and Stockholm, introducing realistic intercontinental delays and variability, and successfully performs a docking maneuver under time-varying random communication latency.
A new framework, Agentic Robot with Tool-use (ART), integrates Visual-Language-Action (VLA) models with agentic tool-use, reducing the complexity of the action solution space through tool-use. ART achieves a 20% higher success rate than mainstream baselines on simulation and real-world tasks, such as pick-and-place in the dark at novel viewpoints. The framework reduces data dependency and improves generalizability across different tasks. ART enables modular tool utilization, efficient training, lightweight deployment, and scalable integration of new tools.