T-Robotics has been selected to lead a KRW 4.5 billion government project with KAIST and a major Korean display manufacturer. The project aims to develop mobile humanoids and test them on manufacturer production lines through December 2030. The goal is to combine environmental perception, decision-making, coordinated movement, and task execution in display manufacturing. This initiative aligns with South Korea's broader humanoid ecosystem plan, focusing on industry-specific robot production systems. The project seeks to gain real-world evidence on robot performance in factory settings to advance industrial automation.
The presentation at ROSCon introduced "The Legible Node," focusing on thin nodes and the Node Definition Language (NoDL) for statically declaring node interfaces. A NoDL document specifies node interactions using sections for parameters, publishers, subscribers, and services/actions. This language allows users to assert node conformity, generate documentation, and map applications statically within ROSGraph. The NoDL project is being developed in the ROSGraph Working Group with milestones focusing on code generation and graph analysis. Interested parties can join the mailing list and resources to contribute to NoDL.
Robotics in manufacturing is evolving beyond simple programmed arms to include advancements in perception, autonomy, and actuation. Demand is increasing across various sectors due to skilled labor shortages, with industrial robot installations doubling in the decade leading up to 2024. Collaborative robots, autonomous mobile robots, and physical AI are enabling more flexible automation solutions. Despite growth, challenges remain in application identification, integration, and ensuring return on investment. Robot order value increased by 6.6% year-over-year in the first half of 2026, with growth seen in non-automotive sectors.
IEEE experts informally discussed the terms "robotics" and "automation" to explore their relationship and significance. Automation focuses on making processes repeatable and efficient, while robotics involves giving automation a physical presence. The distinction is complicated because not all automation requires robots, and not all robots are autonomous. The definitions impact international comparisons, education, and public perception of technological strength. Ultimately, the relevance of the distinction depends on the context and the audience.
NEURA Robotics established NEURA HealthTech, a dedicated business unit for the healthcare and life sciences sectors. The initial application involves an autonomous mobile platform for transporting empty hospital beds, with the MiPA robot intended for patient-room assistance and medication support. This healthcare platform will be offered to partners on a monthly Robotics-as-a-Service basis. MiPA is a wheeled assistant featuring modular hardware and navigation capabilities. Further operational evidence, such as measured staff time saved, is needed to validate the proposed healthcare applications.
IDC estimates nearly 25,000 global humanoid shipments in the first half of 2026, marking a 432.1% year-on-year increase. AGIBOT reported leading the market with over 8,600 shipments, accounting for approximately 35% of the global market. Other estimates, such as Counterpoint's, suggest a similar rapid expansion, though they differ in scale. The majority of shipments still pertain to research, education, performances, and data centers. Shipments alone do not fully indicate autonomous productivity, with IDC forecasting over 750,000 global shipments by 2030.
MindOn demonstrated the Mind-1 robot system focusing on completing useful tasks at a human pace. The demonstration featured dual-arm robots handling parcels and a humanoid collecting laundry in a furnished room. The presentation emphasizes that execution speed is crucial for robots to be effective workers, not just achieving motion. The footage is labeled as real-time and fully autonomous, though the post notes that quantified performance metrics are needed. This development aligns with MindOn's goal of creating a unified intelligence across different robot platforms.
Matic is developing future household robotics, moving beyond floor cleaning toward manipulation capabilities. The company's co-founder states that manipulation is the next phase after developing perception and navigation. A recent video features a stylized humanoid alongside a smaller robot with an arm, hinting at this future vision. Matic emphasizes that floor cleaning is the foundation for broader household robotics. The article notes that the imagery reflects future ambition rather than current shipped products.
Boston Dynamics shared a drone tour of the Atlas production environment to promote hiring opportunities. The video showcases the machining, workshop, and assembly areas involved in manufacturing the humanoid robot. The post specifically directs candidates to roles within the ProtoOps team and other positions related to Atlas development. Openings include manufacturing engineering, software testing, and data operations roles. This visual tour aims to make the workplace tangible for prospective employees interested in building the robot.
AGIBOT deployed 100 X2 humanoid robots across 100 ASD retail stores in China for customer-facing roles in cookware sales. These robots assist shoppers by discussing needs and recommending products while human associates manage promotions and purchases. The deployment includes a scheduled three-day livestream event to showcase store interactions. The interaction involves the robot gathering information and passing complex tasks to human staff for final transactions. The announcement does not provide data on sales impact or the exact rates of human assistance during customer engagement.
Dyna Robotics revealed the Taku robot and Dyna-2.1 system during an autonomous, hour-long laundry demonstration. Taku is a wheeled, semi-humanoid robot designed for tasks requiring reach and mobility in environments like laundry rooms. Dyna-2.1 incorporates whole-body control, a manipulation policy, and a vision-language model for autonomous workflow decisions. The system demonstrates handling a connected sequence of jobs, including recovering from errors. The next planned milestone for the system is customer-site deployment.
The research proposes a framework inspired by passive-dynamic walking to guide learning energy-efficient gait coordination in humanoid locomotion. This method temporarily simulates slope-equivalent conditions during training to encourage economical gait discovery. The framework achieves mechanical cost reduction by leveraging a tilted-gravity field and curriculum-coupled rewards without needing reference trajectories. Experiments on a 29-DoF robot showed a 6.8-15.2% reduction in mechanical cost of transport at commanded speeds of 0.5-2.0 m/s. The approach also extends to omnidirectional locomotion, reducing speed-matched cost of transport by 18.7%.
In-context learning (ICL) allows general-purpose robots to infer task requirements from demonstrations and interactions. ICL directs existing robot competence by keeping neural parameters fixed during deployment. The literature review categorizes ICL interfaces into four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. The analysis examines how these mechanisms preserve learned requirements across changes in objects, environments, and execution conditions. Ultimately, the work connects method design to evaluation to enable compositional task acquisition and physical recursive self-improvement.
SAKI is a framework designed for skill assembly and kinematic imitation from human videos to enable long-horizon mobile manipulation. It connects human-video skill acquisition with cross-demonstration assembly and closed-loop whole-body execution. The framework prepares reusable object-centric skills while allowing transfer paths to adapt across different scenes. It selects and orders skills based on a goal, binds object roles, and carries scene estimates during execution. Real-robot experiments demonstrated skill reuse and composition for complex mobile tasks like tidying.
The MagNav framework provides a dual-core magnetic track guidance system for lighting-invariant navigation in two-wheeled robots. It utilizes a five-channel Hall-effect sensor array to avoid interference from optical sensors, which are susceptible to lighting changes. The control system employs a cascaded PID structure, using an IMU for balancing and magnetic sensor readings for steering. This design maintains accurate tracking even in very bright light conditions, unlike the optical system. Consequently, this method offers a reliable, lighting-independent solution for autonomous navigation in various environments.
Pretrained robotic policies need to rapidly adapt to unobserved physical dynamics during deployment. Existing test-time adaptation methods often fail to use rich geometric and dynamic feedback from physical interaction. SCOUT is a dynamics-aware meta-learning framework that enables policies to adapt by continuously revising internal beliefs about environment dynamics. It couples an action-prediction policy with a forward dynamics model using a shared belief latent space. This approach accelerates online adaptation and improves sim-to-real transfer by updating latent beliefs based on action-outcome mismatches.
KPI is a promptable kernel designed to manage physical interaction for humanoids. It allows a trajectory source to communicate contracts, specifying how the robot should behave upon contact, instead of relying on a fixed controller. The kernel adapts the arms' stiffness, damping, and reference based on tracking error and wrench estimates at the contact rate. This system is demonstrated using an agentic framework where a vision-language agent generates both the reference trajectory and the interaction contract. Examples include instruction-driven winch operation and surface interactions.
SkillWeaver is an agentic framework designed to autonomously generate large-scale robot experience by exploring over Neural Interaction Skills (NIS). These NIS are reusable, parameterized, closed-loop policies that expose learned physical interaction capabilities to a reasoning agent. A VLM agent uses these skills to interact with the environment, observing outcomes and generating reflection to guide further exploration. The framework scales to a large amount of data and results in visuomotor policies that significantly improve generalization in robot tasks. This approach offers a scalable alternative for generating robot data through agentic exploration.
Robot manipulation policies often fail under real-world conditions due to actuator degradation, such as overheating or voltage sags. Telemetry-Aware Action Rectification (TeAR) is introduced as a policy-agnostic method to adapt frozen policies using onboard sensor data. TeAR employs a lightweight Transformer to modify outgoing actions by combining them with live actuator telemetry. Evaluation shows TeAR improves success under heating by 10-15% on a physical arm without requiring on-robot fine-tuning.
A 17-DoF robotic hand was designed using an antagonistic tendon-routing mechanism for bidirectional operation. This mechanism allows active control over flexion/extension and abduction/adduction across all degrees of freedom, enabling grasping on both palmar and dorsal sides. Experimental results show low motor transmission error (3.0%) and good force application (up to 29 N). The hand successfully performed all 33 GRASP Taxonomy grasp types, validating its bidirectional grasping capability. The system demonstrates a compact, human-scale platform for dexterous manipulation.
The work proposes a JEPA-style world model that constrains latent dynamics to a bilinear parameterization for efficient control. This structure is used to enforce action recoverability, which prevents representation collapse by construction. Empirical results show that learning representations with bilinear dynamics significantly reduces planning time while maintaining or improving control accuracy in 2D and 3D tasks. The method is demonstrated to succeed in demanding regimes like longer-horizon planning and real-time control. Ultimately, the approach extends JEPA-style world models beyond short-horizon offline planning.