Regular price ROSCon ticket sales have been extended. This extension is due to support from Destination Toronto. Tickets are now available until Friday, August 28, 2026, at 6:59 AM UTC. This provides more time for ticket purchases.
Humanoid robots are a focus in China's strategy to integrate artificial intelligence into daily life through embodied AI. China is a global leader in manufacturing these robots, accounting for nearly 90% of those delivered worldwide last year. A robot "carnival" in Shanghai showcased various robotic demonstrations, including tasks like making coffee and engaging public displays. These events aim to popularize robotics and address past challenges like safety concerns for humanoids. Chinese firms are actively pushing forward with these developments, attracting public interest.
Pink Robotics, a Darmstadt-based startup, has revealed the P0, a 1.75-meter-tall humanoid robot designed for logistics sorting. The founder criticizes Germany's slow progress in humanoid robotics, positioning the effort as a means to revitalize the German industrial sector. The P0 features 32 degrees of freedom, weighs 82 kg, and has an 8 kg payload capacity. The company is preparing for a logistics pilot with a major supermarket chain to test the robot's capabilities. Pink Robotics is currently raising seed funding to compete with established players like NEURA Robotics and Agile Robots.
AdaPT is a framework that teaches humanoid robots the tennis playing styles of Federer, Nadal, and Djokovic by extracting kinematics directly from broadcast video. It uses a decoupled planning-and-tracking hierarchy enhanced with explicit speed adaptation to compensate for sim-to-real gaps and real-world latency. The system mined 3D body meshes from video to create training data, ensuring the replication of nuanced biomechanics like topspin follow-through. Testing on robots like the Unitree G1 and Dobot Atom showed significant improvements in hit rates and serve success compared to non-adaptive methods. AdaPT demonstrates a method for transferring complex athletic skills from visual data to physical humanoids in real-world settings.
Figure AI demonstrated the autonomous climbing and descending of a 15-foot industrial ladder using its Helix software architecture. This achievement was developed to address a real commercial need for accessing mezzanine work areas in industrial settings. The process required continuous coordination of visual perception, dynamic balance, and contact loading across the entire body. Navigating this task is complex because it necessitates tightly coupling perception, planning, and dynamic control in a continuous manner. This capability is crucial for humanoid robots to achieve general-purpose labor utility in complex, unstructured workspaces.
Chinese humanoid robots shattered human track-and-field records at the 2nd World Humanoid Robot Games. Models like the Tiangong Ultra achieved sub-10-second 100m sprints and sub-40-second 400m runs, surpassing Usain Bolt's benchmarks. Robots also set new marks in the 1500m run and standing high jump. Despite these speed feats, the robots struggled with deceleration and directional stability, resulting in several crashes. Experts suggest the focus should shift from pure speed to achieving practical utility in navigating physical environments.
NEC's corporate venture capital fund invested in US-based robotics startup Dexmate, which develops the VEGA humanoid robot and an integrated physical AI development platform. Dexmate's approach aims to simplify the deployment of physical AI by combining sensors, control systems, and simulation environments. The company seeks to streamline the process of developing and deploying physical AI systems across various industries. NEC plans to integrate Dexmate's technology with its own AI expertise to explore potential applications in Physical AI. This investment supports NEC's open innovation strategy to develop new businesses and technologies.
J&T Global Express added 22 automated sorting systems, bringing the total to 435 across its network. Unmanned delivery vehicles increased by 87% to over 1,900 units by the end of June 2026. The company's goal is to enhance efficiency and reduce costs through outlet automation and intelligent systems. New sorting systems were deployed across Southeast Asia, China, and Other Markets. J&T is also focusing on AI-powered quality inspection and intelligent route planning for its logistics operations.
Selecting roller brushes for solar panel cleaning must balance cleaning action with controlled contact, serviceability, and module limitations. The choice depends on the specific site contamination, cleaning method, and robot interface, not just brush softness. Loose dust is suitable for soft brushing, but hard deposits like bird droppings require more nuanced approaches. Module cleaning guidance and surface protection depend on the entire system, including brush condition and operating conditions. A good roller controls routine soiling while acknowledging when alternative cleaning methods are necessary.
RoboShape introduces information-theoretic compression for point cloud representations to enable privacy-aware robot perception. It uses the Donsker-Varadhan formulation of mutual information to maximize the information between embeddings and object understanding while minimizing the leakage of private attributes. This method results in embeddings that are 87.5% smaller while retaining 98.7% of object classification utility. RoboShape successfully reduces sensitive attribute predictions by 39.3% across indoor LiDAR datasets. The resulting privacy-preserving embeddings are efficient for network transmission and downstream model training.
The system proposes an intelligent companion robot for elderly-care facilities that focuses on natural interaction and safety. It integrates active visual human-following for tracking and a Large Language Model (LLM) for interpreting speech to execute tasks like escorting. A Vision-Language Model (VLM) continuously monitors the environment for fall detection and abnormal posture assessment. This multimodal approach ensures robust human tracking, natural interaction, and critical safety monitoring. Experimental results show the system effectively follows humans while reliably detecting potential falls.
Reliable human tracking for companion robots is achieved using a gimbal-mounted camera that dynamically adjusts its view to keep the target in frame. A continual learning strategy is applied to person re-identification to adapt to changes in appearance and environment during tracking. Experimental results show the system provides stable, real-time re-identification and responsive tracking across various human motions. This method improves user comfort by eliminating the need for wearable tags. Ultimately, the approach enhances the stability and continuity of human motion tracking.
Inertial navigation systems require precise initial alignment for autonomous vehicles, often achieved through Coarse Alignment and Fine Alignment processes. Conventional methods for this alignment suffer from slow convergence and limited bias estimability. This paper introduces an optimized Coarse Alignment method, TRIAD-CBE, and a novel Fine Alignment EKF observation model incorporating Non-Orthonormality constraints. These constraints are derived from TRIAD and directly relate errors to inertial sensor biases. Simulations and experiments show these approaches significantly accelerate alignment and bias estimation while maintaining accuracy.
ODG-NoMaD enhances NoMaD by providing a global sense of direction for exploration in unseen environments. It utilizes an overhead depth camera to create an occupancy map and plan a global path, which is then refined by per-frame traversability maps. A cosine direction cost gradient is injected into denoising steps to rotate trajectories towards this global direction. In simulations, ODG-NoMaD significantly reduces the distance to the target compared to unguided exploration and achieves collision-free paths in all trials.
Dense traffic necessitates an active interaction-aware planning framework due to continuous influence among agents. The proposed method integrates an ego-conditioned generative autoregressive prediction model into Model Predictive Path Integral (MPPI) control. This model generates stochastic, multi-modal predictions of surrounding agents based on ego actions. A nested sampling scheme is used to evaluate expected cost and collision risk under these induced distributions. Closed-loop simulations show improved safety and efficiency compared to traditional planning methods.
World action models (WAMs) jointly denoise future video frames and robot actions, where camera viewpoint change is a major challenge. The study investigates where to apply consistency loss when training WAMs with same-state cross-view image pairs. It demonstrates that imposing consistency on view-covariant coordinates is harmful, shrinking legitimate view-specific content. Selective cross-view consistency (SCVC) constrains only the view-invariant block, eliminating the need for test-time camera information. This method improves closed-loop success in held-out viewpoints without requiring any camera labels or synthesis.
Fixed and portable articulated-arm Coordinate Measuring Machines (CMMs) often differ in their accuracy specifications for the same inspection task. The study investigated whether the difference in measurement results crosses the specified tolerance boundaries, rather than just the magnitude of the difference. Dimensional disagreements were found, with the fixed-CMM being approximately $11.2 \ \mu\text{m}$ higher than the articulated-arm CMM. Form profile differences also showed disagreement relative to the upper limit. The research proposes a tolerance map to identify feature-tolerance combinations where instrument choice can alter the inspection label, necessitating controlled equivalence studies.