RoboCup 2026 is scheduled to take place in Incheon, South Korea, from July 2nd to the 6th. The event encompasses competitions, training sessions, and a symposium for participants. Updates are anticipated regarding both the leagues and the overall competition format. A major focus this year will be placed specifically on the soccer leagues.
AO-ARC is introduced as an anytime multi-robot motion planning (MRMP) method. It achieves initial solution times comparable to existing state-of-the-art feasibility solvers. Crucially, AO-ARC converges faster and more reliably than current anytime MRMP methods as the robot count increases. The technique adapts the AO-x meta-algorithm by iteratively calling ARC on bounded MRMP instances using a makespan cost metric. Theoretical analysis confirms its asymptotic optimality, which was validated through 2D and 3D empirical tests.
Autonomous robotic operations in space require reliable visual references for rendezvous and docking due to environmental challenges. Current single-scale fiducial markers fail during critical close-range proximity phases. This paper introduces AstraTag, a novel marker utilizing a square Spidron pattern with recursive self-similarity for multi-scale detection. Identification employs a 48-bit signature encoded with a GRS code and includes contour-based localization. Critically, AstraTag incorporates a Thin-Plate Spline re-warp fallback to ensure robust performance on curved spacecraft surfaces compared to existing benchmarks.
SceneBot is a new motion-tracking framework designed to overcome limitations in current humanoid reinforcement learning when performing contact-rich tasks. It unifies control for free-space locomotion, terrain traversal, and whole-body manipulation by conditioning a single policy on reference motions and per-link contact labels. To handle data scarcity, the authors propose a hindsight scene reconstruction method that infers interaction graphs from human motion. Trained on 7.5 hours of this reconstructed data, SceneBot generalizes effectively to novel environments and movements. The framework demonstrates the ability to seamlessly execute complex tasks requiring both free movement and object interaction.
Quadrotors are valuable for time-critical transport like search-and-rescue missions, but precisely releasing suspended payloads remains an underexplored area. Existing model-based methods struggle with rope flexibility and tracking errors due to overly conservative constraints. Researchers propose a novel hybrid simulation framework coupling a high-fidelity quadrotor model with a physics solver for accurate rope-payload force exchange. A deep reinforcement learning policy trained in this environment achieves agile, accurate payload throws, outperforming baselines by reducing landing error up to 50%. Furthermore, the system can be trained using only visual observations, demonstrating comparable accuracy to state estimation methods.
Developing reliable techniques for locating radiation sources using robots is crucial for enhancing safety and efficiency. Existing approaches often require robots to approach the source, which poses risks of radiation damage. This work introduces an automation framework utilizing a physics-informed machine learning (PIML) model for precise source localization along any measurement path in unknown environments. The method incorporates physics-inspired tensors to manage attenuated gamma-ray flux signals from various obstacles while computing multiple models concurrently for improved robustness. Validation was performed through high-fidelity simulations and physical experiments, advancing robot radiation perception toward spatial intelligence.
Presents Parallel ARC (P-ARC), a parallel extension of the Adaptive Robot Coordination (ARC) method for multi-robot motion planning (MRMP). P-ARC achieves parallelism across the three core stages of ARC: initial solutions, conflict detection, and conflict resolution by exploiting problem independence. A hybrid strategy, OR-P-ARC, is introduced using an OR-parallel multi-start approach with both ARC and P-ARC. Evaluations utilized 2D mobile and planar manipulator scenarios involving up to 128 robots to test parallel efficiency. The research demonstrates planning time speedups approaching four times over the sequential version for large Panda multi-manipulator teams on 16 CPU cores.
Vision-Language-Action (VLA) models struggle with spatial and task changes when deployed. Previous methods relying on language or visual prompting fail to solve this brittleness effectively. The research proposes a novel, model-agnostic module that represents grounding signals using 3D points. This 3D embedding is computed relative to the gripper and injected directly into the action head via adaptive layer normalization. Implementation on the LIBERO-PRO benchmark significantly boosted success rates for both task and position perturbations, confirming that 3D injection into the action head is key to generalization.
Developed Composite Whole-Body Imitation (CWI) to address challenges in coordinating stable locomotion and versatile manipulation for humanoid robots. CWI decouples upper-body manipulation data from lower-body locomotion control, allowing the use of diverse MoCap datasets while ensuring stable movement. Lower-body dynamics are maintained using an Adversarial Motion Prior guided by dual discriminators trained on curated walking and squatting clips. The framework employs a multi-critic architecture to balance conflicting objectives among motion style, locomotion, and manipulation. Testing on simulation and a physical LimX Oli robot demonstrated competitive performance with robust whole-body coordination from minimal input commands.
World-action models often struggle with long-horizon tasks because they rely too heavily on short-term history. DiM-WAM addresses this by introducing a memory-augmented framework to incorporate multi-scale historical context, future dynamics, and global task progress. The model extracts compact visual event information into multiple merging memory banks. It uses these banked memories to condition video and action denoising for better long-term awareness. Evaluation on RMBench showed DiM-WAM boosting average success from 28.4% to 69.8%, significantly outperforming existing baselines on both simulated and real-world tasks.