Steve Crowe, Mike Oitzman, and Sarah Wynn reviewed Automate 2026. They analyzed significant trends observed at the event. Key areas of focus included physical AI advancements. Furthermore, humanoid development was a major topic discussed. Finally, software orchestration practices were also reviewed by the editors.
Integrating mechanical positioning is increasingly recognized for enhancing a machine's overall performance in complex tasks. This technology significantly impacts key operational metrics, including the robot's mobility, range of motion, and speed. Combining advanced robotic dexterity with precise mechanical positioning enables more sophisticated handling of assembly operations. Such integration allows for greater capability in intricate manufacturing processes. Experts advocate for this combination to optimize complex assembly workflows.
Luxonis successfully closed a Series A funding round. Funds will be used to scale the production of its OAK camera. The company also intends to refine its platform for broader applications. These efforts will specifically target robotics and industrial automation sectors. This strategic investment aims to advance Luxonis's physical AI perception layer capabilities.
The 2026 IEEE International Conference on Robotics & Automation (ICRA) convened experts in robotics over one week. The event took place in Vienna, specifically at the Messe Wien. Attendees gathered to present cutting-edge technological advancements across the field. A major focus was on fostering new professional collaborations among researchers and industry leaders. The conference served as a vital platform for exchanging the latest ideas within robotics.
RoboCup 2026 began today in Incheon, South Korea, with league competitions scheduled until July 5th. This year marks an exciting period for RoboCup due to updates to both leagues and competition formats. A major change is the increased emphasis on humanoid robots within the soccer leagues. The event kicked off the competitive action for the tournament's participants.
Embodied artificial intelligence is rapidly transforming robotics, driving development towards general-purpose machines capable of diverse real-world tasks. Companies are competing to achieve this versatility beyond single-application programming. Shenzhen-based X Square Robot is a notable leader in this field. The company recently secured funding through four consecutive financing rounds. This effort underscores the industry's push toward comprehensive robotic intelligence solutions.
AGIBOT showcased its A3 humanoid robot in Europe during the UK AGIBOT Partner Conference 2026 in London. The event signifies a major component of the company's European expansion plan. Furthermore, AGIBOT introduced a Robot-as-a-Service model specifically for the United Kingdom market. These developments highlight the company's focus on commercializing humanoid robotics. Their strategy integrates technological innovation with practical deployment and local business collaborations.
Autonomous UAV landing on wave-disturbed marine platforms is challenging due to unpredictable motion and attitude variations. Existing methods struggle with either complex online optimization or limited interpretability. WaveLander introduces a novel, hierarchical reinforcement learning framework to address this difficulty. The approach separates the vertical landing decision from low-level flight stabilization control. Simulation results confirm that WaveLander achieves robust performance across randomized wave disturbances while generalizing effectively.
ROS 2 utilizes middleware like DDS and Zenoh for distributed robotic communication, yet these systems face structural limitations in dynamic wireless settings. This paper proposes examining ROS 2's architectural limits using three dimensions: Space, Time, and State. Time relates to temporal predictability essential for control loops, while Space addresses abstracting physical topology for modularity. State management ensures contextual continuity despite node changes and connection interruptions. The work reveals structural trade-offs where achieving spatial abstraction or state continuity can compromise time-critical communication under wireless constraints.
Vision-Language-Action (VLA) models' real-world use is hindered by insufficient safety measures beyond merely preventing immediate collisions. This paper introduces a neuro-symbolic guidance mechanism for flow matching based VLAs to enable predictive collision avoidance. The method frames safety enforcement as a minimum-norm constrained optimization problem applied during the trajectory denoising process. By analyzing and correcting predicted trajectories iteratively, it anticipates potential future collisions rather than just reacting to them. On the SafeLIBERO benchmark, this approach showed significant improvements, achieving 82.8% collision avoidance and better task success rates compared to single-step methods.
Introduces CommonRoad-Game, a lightweight human-in-the-loop simulation framework designed for testing autonomous driving motion planning algorithms with human participation. It addresses limitations of existing simulators by offering real-time interaction capabilities and being suitable for rapid prototyping. The framework features a multi-threaded architecture ensuring deterministic synchronization between simulated and real time. CommonRoad-Game integrates seamlessly with the CommonRoad platform, supporting scalable multi-agent simulations. Additionally, it includes a module to record driving logs, enabling the creation of diverse and reproducible test cases from human interactions.
Challenges in sim2real robot policy learning stem from mismatches between simulation and reality, which current methods address piecemeal. This work proposes BIFROST to exploit the shared structural basis underpinning task equivalence across domains. It introduces a shared history encoder trained using a cross-domain bisimulation objective on paired data. This objective ensures that observation-action sequences resulting in equivalent long-term behavior map close together in latent space, irrespective of domain differences. Empirical results confirm BIFROST enables zero-shot transfer for visual navigation and contact-rich manipulation where other methods fail due to combined visual or dynamics gaps.
Proposes SE(2) Navigation Mesh (SE(2) NavMesh), a polygonal representation designed to capture yaw-dependent traversability for ground robots, overcoming limitations of traditional point clouds and meshes. It models traversable areas using footprint masks and builds a graph with explicit translational and rotational connectivity across different yaw layers. A novel pathfinding strategy, A*-String Pulling-A* (ASA), utilizes this mesh to hierarchically optimize robot position and heading. The method also includes an online procedure for incrementally updating the SE(2) NavMesh from real-time point cloud streams. Simulations show the SE(2) NavMesh captures significantly more traversable area, and its integration with ASA outperforms existing baselines in complex scenarios tested on physical hardware.
Proposes a reconfigurable rocker-bogie mechanism enabling efficient turning and high step-climbing with few actuators. The system can switch between four-wheel and six-wheel configurations by actively swinging bogies using motors at the joints. Omnidirectional wheels on the rockers allow for smooth, zero-radius turning modeled by differential drive in the four-wheel setup. Experimental results show the robot achieves this turning five times faster than conventional mechanisms, using only 17% of the average wheel torque. Furthermore, the prototype successfully climbed a 40 cm step in an average time of 6.4 seconds.
A novel layered planning and control framework using multi-rate nonlinear model predictive control (MR-NMPC) allows quadrupedal robots to achieve hybrid bipedal locomotion while supported by walls. This system dynamically plans both discrete contact points and continuous Center of Mass/orientation trajectories using a single rigid body dynamics model. The incorporation of explicit contact-point planning enhances dynamic stability over traditional heuristic foot placement methods. At the lower level, a nonlinear whole-body controller uses virtual constraints to track the MR-NMPC references across full-order dynamics. Simulations confirmed the robust wall-assisted locomotion on rough terrains, achieving a 2.9 times higher success rate than conventional MPC approaches.
Addresses challenges in real-time collision avoidance for unmanned surface vehicles (USVs) against both fixed and moving obstacles. Proposes a system integrating a novel global path planner with a local path planner for robust navigation. The global planner combines Grassfire, a modified Grassfire, and an intuitive Probabilistic Roadmap variant to map obstacle-free paths from start to goal. For dynamic avoidance, the local planner uses advanced decision logic based on observed relative obstacle movement. Simulations validate this approach, showing superior performance compared to established methods like D*, particularly in achieving systematic routing behind moving objects.