Two teams of 11 humanoids competed in the first 11 vs 11 soccer game at RoboCup 2026. The match featured B-Human from Bremen, Germany, against HTWK Robots from Leipzig, Germany. Both teams utilized machines designed by Booster Robotics. The event highlighted advancements in humanoid robotics, noting improved teamwork and AI-driven intelligence. This competition represents progress toward the long-term goal of autonomous robot achievement.
AGIBOT won the overall medal table at the 2nd World Humanoid Robot Games in Beijing, securing 46 total medals. This victory was achieved through strong performance in dexterous manipulation and scenario-based tasks, with its OmniHand earning seven gold medals. AGIBOT utilized mass-produced hardware like the G2, A3, and X2, reflecting an industry shift towards real-world utility. TianGong Robotics narrowly finished second with 45 medals, dominating speed and jumping events. The results highlight a developing trend where multi-domain competence is becoming as important as peak athletic performance in humanoid robotics.
A demonstration was performed using a global-shutter camera and an IMU streamed over a single GMSL2 coax cable. This data was mapped live in ROS 2 using a Jetson platform and OpenVINS. The system utilized an NXS sensor module featuring a Sony IMX900 global shutter and an IAM-20680 IMU. The setup involved real-time estimation without any wheels, markers, or prior map. The test resulted in a return-to-start error of 42 cm after a 25-meter loop.
Automotive glass manufacturing is difficult to automate due to the material's tendency to shatter and the precision required for downstream processes. Modern automation uses six-axis articulated robots combined with vision systems to accurately pick and place irregularly shaped glass blanks. The shift to electric vehicles is increasing demands for higher precision and consistent quality in glass components. Successful automation requires assessing volume consistency, geometry stability, breakage costs, and downstream process tolerances before investment. Future advancements will focus on predictive breakage models to move from reactive handling to proactive quality control.
Kodiak AI and AMD have collaborated to use AMD EPYC Series processors for Kodiak's seventh-generation autonomous truck platform. This partnership allows AMD to provide the necessary compute power for Kodiak's driverless trucking technology. The AMD CPUs handle the aggregation and preprocessing of massive sensor data from LiDARs, cameras, and radars for safe operation. This enhanced processing capability supports latency-sensitive tasks like sensor processing, localization, and path planning within the Kodiak Driver. This collaboration aims to advance driverless trucking from innovation to scalable commercialization.
GUSS Automation, a John Deere subsidiary, plans to integrate Ouster's Rev8 OS0 native color digital lidar sensors into its autonomous orchard machine fleet. This integration aims to enhance machine capabilities by incorporating advanced sensing technology for navigation in complex outdoor environments. Traditional GPS often lacks the necessary centimeter-level precision for safe navigation in dense orchards and vineyards. The lidar technology provides dense 3D point cloud data, enabling superior navigation and obstacle detection under various conditions. This development empowers growers with increased precision and efficiency for autonomous agricultural operations.
A constraint-aware physics-informed neural network (PINN) is proposed for the static shape estimation of co-manipulative continuum robots (CCRs). The method incorporates residuals for static equilibrium and geometric loop-closure constraints based on the geometric variable strain formulation. Simulation results show the PINN significantly reduces errors compared to purely data-driven models when trained with limited and noisy data. The PINN achieved high accuracy with fast inference times, demonstrating physically consistent and computationally efficient shape estimation. Experimental fine-tuning further improved marker RMSE and the R-squared value of the estimation.
The paper develops a propagation-based kinodynamic planner with near-optimal guarantees for state-to-state boundary value problems in nonlinear systems. It proposes a Dispersive Forward Tree search method using locally dispersive control commands to find near-optimal trajectories. The method includes a cost-conditioned dominance pruning procedure to maintain near-optimality while keeping the search tree size polynomial in the horizon. The resulting Dispersive Forward Tree search (DFT*) is implemented as a breadth-first expansion suitable for parallel hardware. DFT* demonstrates competitive and improved solution quality over existing planners on embedded processors.
The research benchmarks neural 3D reconstruction methods like NeRF and 3D Gaussian Splatting across various compute platforms for autonomous laboratory robots. The study evaluates Meta's SAM3D single-image reconstruction on these same platforms to assess performance gaps. Results indicate that Gaussian Splatting offers better rendering quality than NeRF at higher computational costs. Onboard compute is generally insufficient for full per-scene optimization at interactive rates. Consequently, a tiered pipeline is suggested where lightweight methods handle real-time loops and heavier methods are applied selectively.
Zero-shot object navigation methods struggle with performance under real-time execution constraints due to inference latency. The proposed RTNav architecture addresses this by explicitly considering inference latency, asynchronous stepping, and compute bounds. This approach treats these factors as explicit design considerations rather than ignoring them in sequential execution models. Evaluation on real-time variants of HM3D models showed RTNav improved success rates by up to 11% and success weighted by completion time by up to 5.1 points. RTNav demonstrates a method for more efficient, real-time navigation in unknown environments.
The research addresses the lack of a reliable methodology for unassisted bipedal locomotion on the Poppy Humanoid. The authors propose a closed-loop walking controller utilizing the Linear-Quadratic Regulator (LQR) framework for trajectory tracking. They learn a quadratic cost function for the LQR controller using data from open-loop playback of a nominal walking trajectory. This learned cost function significantly improves the reliability of the motion. Empirical validation showed statistically significant improvements in walking performance compared to open-loop playback.
To prevent catastrophic forgetting in continual robot learning, the paper introduces "Memory Anchors" derived from past experiences. These anchors are identified in regions where new task representations overlap with old ones despite conflicting action requirements. Rehearsing data in these anchor regions is crucial for preventing the overwriting of prior knowledge. Including these anchors in the replay buffer significantly reduces high-conflict task forgetting. This approach enables successful continual learning across multiple task sequences on real robots.
This work introduces Configured Failure Trapping, a novel backdoor attack task for Vision-Language-Action (VLA) models that requires controlling specific failure modes. It necessitates attackers to control how the robot fails, making the attack more challenging and harder to detect than previous methods. The authors developed a data engine and measurement suite to support this task, creating benchmarks like Trap-LIBERO and Trap-RoboTwin. TrapVLA is proposed to learn trigger-induced action residuals to successfully steer the policy into the configured failure behaviors. Experiments confirm that TrapVLA effectively injects these failure modes while largely maintaining performance on clean data.
SOLO is a framework designed for stable, long-horizon, perceptive humanoid locomotion across various terrains. It addresses fragility in humanoid policies by managing dense terrain reconstruction and temporal credit assignment errors. The Query Reconstructor (QR) uses Fourier-encoded queries to preserve sharp terrain boundaries from depth-proprioception data. Trajectory-Aware MSE Distillation enhances learning by incorporating next-state disagreement into the reward signal. SOLO achieves high success rates on stress-test terrains and can complete autonomous navigation routes using limited sensory input.
Soft robotic grippers struggle with force decay due to viscoelastic stress relaxation during holding. This research introduces a framework combining structure-perception and learning for relaxation-aware sensing. The system uses onboard vision and infrared thermography to track soft gripper deformation and temperature in real time. A temperature-coupled viscoelastic force representation and a physics-informed learning model are used to compensate for force decay. Experiments demonstrated this method maintains desired grasp force with high accuracy, outperforming existing baselines.