NVIDIA Cosmos has expanded its platform with new tools, checkpoints, and blueprints to help developers build, train, evaluate, and deploy world models for physical AI systems. The Cosmos Data Factory facilitates large-scale data curation, processing, and synthetic data generation for training multimodal datasets. The Cosmos Framework provides agent skills for fine-tuning open models and optimizing them through techniques like distillation and quantization. New Cosmos 3 Nano models and NIM microservices enable faster testing and deployment of autonomous policies and video analysis. Overall, the platform offers a modular approach for the entire world model development lifecycle.
European retailers are leveraging NVIDIA's open models and physical intelligence to transform commerce and operations. Companies like Shopify and Zooplus are using these technologies to advance digital discovery and product search with significant cost reductions and improved accuracy. Physical AI enables simulations of warehouse operations and virtual try-ons using digital twins, as demonstrated by Otto Group and ALLSIDES. This integration allows for optimizing supply chains, generating accurate product representations, and accelerating research in areas like material science. Ultimately, NVIDIA solutions are enabling a shift from digital discovery to tangible physical operations in the retail sector.
Day 2 of RoboBusiness 2026 begins with a "State of Humanoids" panel featuring industry leaders discussing humanoid capabilities and challenges. The day includes an Exhibitor Hall open for demos and a final keynote on robotic character creation. Morning breakout sessions cover topics like egocentric robot learning, China's AI ecosystem, and physical AI failures in real-world applications. Afternoon sessions address topics such as scaling physical AI, policy discussions, and real-time robot troubleshooting. Overall, the day provides in-depth technical and business insights into the robotics industry.
Intramotev will deploy its TugVolt autonomous freight railcars on Missouri’s Columbia Terminal Railroad (COLT) as part of a $16.2 million federal grant. This project involves testing autonomous operations on COLT's 21.4-mile line between Centralia and Columbia, Missouri. The deployment will begin with remote switching at the Columbia terminal, progressing to full autonomous freight service. The grant also covers necessary track and infrastructure upgrades for the railroad. The resulting safety record will serve as a template for other short lines adopting autonomous railcars.
Automated warehouse systems necessitate specialized safety checking of racking to ensure structural soundness and compliance. Automation removes manual checks, shifting the focus to inspecting complex issues like rail wear and fatigue in high-bay and dynamic systems. Proper inspection involves a structured engineering assessment by a competent inspector, resulting in documented risk classifications. While monitoring tools can flag anomalies, a formal inspection provides the necessary qualified judgment and legal sign-off. Consequently, SEMA-approved inspections remain crucial for ensuring the safety and continued operation of automated storage infrastructure.
Tesla began producing its Optimus humanoid robot in Fremont, California, utilizing factory space previously used for electric vehicle manufacturing. This production marks the first batch of quality bots, though CEO Elon Musk anticipates a slow ramp-up due to the product's novelty. Despite this development, the company experienced a 43% drop in profit for the third quarter due to heavy investment in Optimus, AI, and autonomous driving. Revenue remained flat, driven largely by EV sales, while expenses for these projects caused profits to decline significantly. Net income fell to $787 million, a substantial decrease from the previous year.
Voxel is a humanoid home robot developed by Tuesday Labs, featuring a boxy yellow head, colorful panels, and Z-shaped legs, standing 150 cm tall and weighing 45 kg. It is powered by electric actuators and features stereo RGB cameras for perception and seven degrees of freedom in each arm. Demonstrated capabilities include obstacle avoidance and sorting goods, establishing a foundation for domestic organization. As a prototype, shipping is planned for 2028, and the robot is aimed at bringing physical robots into everyday living spaces.
AGCO is piloting Agility Robotics’ Digit humanoid at its Hesston, Kansas, plant starting in 2027 for parts handling. The robot will sort parts and place totes into storage within the existing facility. The goal is to automate repetitive, physically demanding tasks to reduce employee strain and improve efficiency without major infrastructure changes. This pilot tests the robot's ability to perform useful work in human-centric factory workflows. The project involves AGCO joining other customers like Toyota, GXO, and Schaeffler.
Accurate ground plane normal estimation is crucial for odometry in spherical mobile mapping systems due to rapid rotations and limited coverage. Two RANSAC-based methods, single-scan estimation and multi-step validation, are proposed for this task. LiDAR-Inertial Odometry (LIO) estimation, utilizing a global 3D map, is shown to significantly outperform single-scan methods in accuracy and robustness. Integrating these ground normal estimates into the motion model enhances pose estimation, aiding autonomous navigation. This work contributes to improving motion estimation in challenging mobile mapping environments.
The research explores coupling visual and motor remapping for first-person humanoid teleoperation to expand operator access. It introduces a system that warps the workspace by pairing a panoramic azimuth warp with an inverse upper-limb yaw transform. A working prototype implements Global Zoom (GZ) and Focus+Context (F+C) functionalities for this remapping. Within-subject testing showed that Global Zoom resulted in the highest task completion count, while Focus+Context was preferred by some participants. This study demonstrates that different visual-motor mappings yield varied performance outcomes in humanoid teleoperation.
The framework addresses human-robot interaction (HRI) by recognizing interaction modes (active guiding or passive following) without predefined thresholds. A meta-learning neural network is developed to recognize human intention by incorporating force-dependent parameters from the controller. An adaptive meta-learning controller is introduced to compensate for human disturbances and adapt to varying interaction forces online. A multiple-trajectory meta-learning algorithm jointly learns environment representations and force parameters for handling unmodeled dynamics. This approach results in robust mode recognition and adaptive control for safe physical HRI.
A novel framework is presented for physical Human-Robot Interaction (pHRI) that integrates a wrench observer with a hybrid pose and force controller. A neural network is used to enhance observer performance by compensating for friction and unmodeled dynamics. Task-null-space decomposition is applied to separate joint velocities and mitigate human disturbance effects on the observer. This method achieves robust wrench estimation and reliable pose/force tracking without requiring force/torque sensors. The approach is effective for pHRI on robotic systems lacking force/torque sensors.
VeriCalib addresses the limitation of traditional LiDAR-camera calibration by formulating it as a verifiable estimation problem. It introduces a two-level extrinsic compensation mechanism to refine the calibration, handling rigid errors at the residual level and rotation/translation at the parameter level. Calibration is assessed using a structured covariance that incorporates uncertainty propagation and known-geometry priors. The method further employs in-view and cross-view checks with thresholds to validate the results under unknown noise. Ultimately, this process yields a calibration state that guides data collection and determines the reliability of the estimated extrinsic parameters.
T-WLA introduces a world-language-action architecture to bridge slow visual processing and fast tactile feedback in contact-rich manipulation. It utilizes shared visual memory to guide both action generation and tactile correction. A visual branch integrates visual information and history into memory to condition base action generation. This memory is cached for a tactile editor that predicts residual corrections based on active actions and incoming touch. Evaluation shows T-WLA reduces finger-joint MPJPE by 30% and achieves a 50.0% success rate on real-robot manipulation tasks.
Real2Sem2Real creates thousands of robot demonstrations from a single human RGB video using 3D semantic augmentations. The method tracks object poses with DINOv3-distilled Gaussian splats to estimate metric depth. It aligns digitally generated object cousins with the demonstration based on semantic information. This process maps the human hand to feasible robot grasps, generating diverse demonstrations for robot learning. Policies trained on this data match or exceed those from teleoperation across various object instances.
A memory framework is proposed where robot episodes are encoded as separable objects using neural implicit representations. These representations are fitted to episode video from a shared initialization, and the resulting weight delta serves as the memory, maintaining a fixed size regardless of episode length. A learned projector maps this weight delta into token embeddings to condition language models for question answering and command regression. This method allows for scalable streaming memory by decoupling memory cost from the time horizon. Experiments show this approach yields strong performance in language model tasks.
Trajectory demonstrations often fail to specify the underlying reasons for robot behavior, leading to inaccurate internal representations. The proposed Flexible Feature Editing Pipeline (FLEX) uses language models to iteratively refine representations of task features through three operations: adding missing features, generalizing existing ones, and masking irrelevant ones. This method addresses limitations of existing representation editing by allowing for more nuanced feature specification. Evaluation on various tasks demonstrated that all three operations are necessary for interactive scenarios. FLEX consistently outperformed other feature editing approaches while maintaining compact and transferable representations.