Robotics News 2026

Robotics in 2026 is no longer defined by one spectacular humanoid demo. The field is being shaped by four connected shifts: robots are moving into real production tasks; embodied-AI models are becoming better at planning and self-correction; edge hardware and safety stacks are maturing; and China, the United States, Korea, Japan, and Europe are all building industrial ecosystems around physical AI.

On July 15, NVIDIA introduced the Jetson T3000 and T2000, smaller Thor-based modules aimed at mainstream robots and edge-AI systems. T3000 is designed for heavier multimodal and robot-policy workloads, while T2000 targets mobile robots, manipulators, and visual-AI agents. NVIDIA also introduced Cosmos 3 Edge, a 4-billion-parameter world model intended to run on Thor hardware for on-device perception, reasoning, and action prediction. The new modules are scheduled for availability in the first quarter of 2027, so this is an infrastructure roadmap rather than evidence of broad deployment today.

On May 31, NVIDIA announced the Isaac GR00T Reference Humanoid Robot, an open research design combining a Unitree H2 Plus body, Sharpa five-fingered hands, Jetson Thor compute, and the Isaac GR00T development platform. Ai2, ETH Zurich, Stanford, and UC San Diego were named among the first research users, and availability through Unitree is planned for late 2026.

On June 22, NVIDIA introduced Halos for Robotics, a full-stack safety system spanning IGX Thor, sensor connectivity, Halos OS, outside-in safety blueprints, and an inspection lab. Agility Robotics is incorporating parts of the system into Digit, making safety architecture—not only model capability—a central competitive layer.

On June 24, Agility Robotics announced a definitive business-combination agreement with Churchill Capital Corp XI. If the transaction closes, the combined company is expected to list under AGLT. The announcement matters because it links humanoid-robot commercialization to production scale, customer orders, and operating data—not only laboratory demos.

The 2026 World Robot Conference opened in Beijing on August 19 with more than 300 exhibitors, over 2,000 expected exhibits, and more than 150 expected product debuts. On the same day, Unitree began trading on Shanghai’s STAR Market after raising about $904 million, according to the Associated Press. Together, the events show how quickly humanoid robotics is becoming a manufacturing, supply-chain, and capital-markets story—not only a research race.

On August 14, LG announced that it is developing a bipedal humanoid using NVIDIA Isaac GR00T, Jetson Thor, and Halos for Robotics, with a public unveiling targeted for the first quarter of 2027. LG also plans to validate its wheeled CLOiD robots on a Tennessee washing-machine production line in 2026 and use the resulting data to improve its own robot foundation model. The important point is the loop between hardware, factory trials, data collection, model training, and safety—not simply the promise of another humanoid.

On July 16, Hyundai Motor Group said its shareholders were pursuing the acquisition of SoftBank’s remaining stake in Boston Dynamics, subject to approvals and settlement procedures. Hyundai described an end-to-end robotics value chain that combines Boston Dynamics’ AI and robot engineering with Hyundai’s manufacturing and global operations. The group currently plans to introduce Atlas at Hyundai Motor Group Metaplant America in 2028 for parts-sequencing work and potentially expand into component assembly by 2030, subject to validation and operational readiness.

The industrial story is therefore moving at different speeds: some robots are already performing limited production work, while others remain pilots, planned deployments, or research platforms. The distinction matters when comparing polished demonstrations with repeatable factory performance.

Hexagon Robotics’ wheeled humanoid AEON began performing production tasks at BMW Group Plant Leipzig while being trained for future applications. In a June 2026 update, the companies said the next phase would focus on autonomy, robustness, and operating performance, with battery-assembly activities expected to progress from the Innovation Garage into the factory and production deployment targeted by the end of 2026. That makes AEON a useful test of whether a humanoid form can deliver dependable value inside an existing automotive workflow.

Google released Gemini Robotics ER 2 on July 30. The model is designed to sit above lower-level vision-language-action models and robot APIs: it watches continuous video, tracks progress, plans multi-step tasks, self-corrects, calls tools, and coordinates different robots. Google also added safety evaluations for instruction following and human proximity. ER 2 is publicly available through the Gemini API and Google AI Studio, while enterprise access remains in private preview.

On August 10, Dyna Robotics introduced DYNA-2, a world-action model pretrained on more than one million hours of egocentric human video. The company reports that the model transfers across robot arms, humanoid prototypes, and dexterous hands, then adapts with relatively small amounts of local robot data. The broader significance is the proposed escape from robotics’ data bottleneck: abundant human video can supply physical priors, while scarce teleoperation data is reserved for embodiment-specific fine-tuning. The performance figures are company-reported and should be treated as results that still need broad independent replication.

Nvidia found a near-perfect log-linear scaling law (R² = 0.998) between human data volume and action loss, directly predicting real-robot success. They created an EgoScale framework, which helped a humanoid with 22-DoF hands learn to assemble model cars, operate syringes, sort cards and fold shirts from 20,000+ hours of egocentric human video – no robot in the loop during pretraining. With just 4 hours of robot play data, the policy achieves 54% gains over training from scratch and even transfers to a 7-DoF Unitree G1 with 30%+ improvement. This means that scaling human motion may be the most practical path to robot dexterity. Explore the EgoScale paper and demos →

Physical Intelligence is building a shared “intelligence layer” for robots – like APIs, but for physical action. They teamed up with Weave Robotics and Ultra Robotics to run π0.6 in real deployments: folding laundry at Sea Breeze Cleaners with 92% autonomy and packaging warehouse orders at 165 items/hour with minimal interventions, cutting interventions by up to 50% and improving throughput with each generation. So instead of engineering full stacks from scratch, companies can plug their hardware into π0–π0.6 models and benefit from shared foundation models to scale real-world deployments. Read how π0.6 handles laundry folding and warehouse packaging in real deployments →

Cornell and Stanford Universities proposed a way to teach robots how to use tools without hand-crafting every task. Instead of training on one tool at a time, SimToolReal trains a single reinforcement learning policy in simulation on lots of generated tool-like shapes. Now one policy can use new real-world tools zero-shot. The results are impressive. Read the SimToolReal paper and try the live browser demo →

In a hands-on project, Paul Ruiz combined a vision-language-action model for pick-and-place control with Gemini 3 Flash for rules and game-state tracking. Trained on 400 teleoperated episodes, the robot arm learned to play the children’s board game First Orchard. The demo is smaller than the industrial systems above, but it illustrates the same architecture now becoming common in robotics: a high-level reasoning model coordinates a lower-level action policy. Read the project write-up →

KV-Tracker makes advanced multi-view 3D vision models more practical for robotics. It enables a robot to track objects or entire scenes in real time using a monocular RGB camera. By caching key visual information, the authors report up to a 15× speedup and roughly 27 frames per second, supporting 6-DoF pose tracking and online 3D reconstruction for manipulation, navigation, and interaction. Explore the KV-Tracker paper and demos →

Menlo Research’s Asimov project is developing a bipedal humanoid around parts and manufacturing methods that smaller teams can access. The currently published Asimov v0 work focuses on an open bipedal leg design using off-the-shelf motors, printable components, 12 total degrees of freedom, articulated toes, and a specialized ankle mechanism. The larger v1 body, simulation files, and actuator list remain part of the project’s roadmap, so readers should distinguish the available repository from planned releases. Explore Asimov on GitHub →

ElRobot is a low-cost, fully 3D-printed robotic arm designed for physical-AI research and imitation learning. The project lists a cost of about $220 per arm, 7+1 degrees of freedom, and a 430 mm reach. It uses off-the-shelf servos and printable parts, supports leader–follower teleoperation, and includes camera mounts, making it a practical platform for small research teams and builders. Explore ElRobot on GitHub →

The biggest developments are the move from demonstrations to factory pilots and production tasks; stronger embodied-reasoning and world-action models; larger human-video and robot-data pipelines; more on-device AI compute; and growing emphasis on safety, reliability, certification, supply chains, and deployment economics.

A humanoid robot has a body plan designed around human environments, usually with a torso, arms, legs or a wheeled lower body, sensors, and manipulators. The form helps it use doors, tools, shelves, workstations, and spaces that were designed for people.

Prominent developers include Boston Dynamics, Agility Robotics, Tesla, Figure, Apptronik, Unitree, AGIBOT, UBTECH, Fourier Intelligence, Sanctuary AI, NEURA Robotics, and Hexagon Robotics. Large manufacturers including Hyundai, LG, and BMW are also building, funding, or deploying humanoid systems with specialist partners.

Robotics foundation models connect language, vision, video, sensor history, and actions. In current systems, a high-level model such as Gemini Robotics ER 2 may plan and monitor a workflow, while a lower-level vision-language-action policy or robot controller executes motion. World models such as DYNA-2 and Cosmos 3 Edge aim to improve physical prediction and action generation, but robot-specific data and safety controls are still required.

Yes, but mostly in tightly scoped pilots and limited deployments. Robots such as Digit and AEON are being tested or used for material handling and production tasks, while Hyundai’s Atlas deployment is planned for 2028. These programs are evidence of real industrial progress, but they are not yet broad replacement of human labor across factories.