The Physical AI Technology Primer.

Seven layers

Body, sensing, compute, control, models, simulation and data, then fleet operations. Every robot is a stack of these.

Three speeds of control

An AI brain at about 10 to 30 Hz, whole-body control at tens to hundreds of Hz, motor loops at thousands of Hz.

Plumbing is shared

ROS 2 and EtherCAT are common standards. Makers compete on mechanics, data and the trained model.

Trained in simulation, finished on site

Simulation does the bulk. A staged live period with human rescues fixes what it missed.

VLA models are the new brain

Camera images plus a spoken instruction in, joint actions out. Open and closed families both exist.

Contact is the hard part

Gravity is easy to simulate. Friction, soft objects and clutter drive the sim-to-real gap.

Real data is scarce

NVIDIA suggests about 100 to 500 good demonstrations to fine-tune a VLA on a new fine task.

Data ownership shapes design

Large customers want training to run in their own cloud account, with a thin local fine-tune.

Safety rules are moving

ISO 10218 was reissued in 2025, ISO 25785-1 is in draft, EU Machinery Regulation applies 20 Jan 2027.

The business is service

RaaS, uptime and remote operators decide returns. GXO said in Aug 2026 that 45 humanoid pilots had not yet paid back.

Sense, decide, act

• AI that perceives and acts on the physical world: robots, drones, vehicles and smart machines.
• A closed loop: sensors feed a model, the model drives actuators, the world changes, sensors see it again.
• Humanoids, arms, wheeled robots and quadrupeds all share this loop.

Why it is harder than chatbots

• A wrong answer costs damaged goods or injury, not a retry.
• There is no internet of robot actions to learn from, so data is scarce.
• Latency, battery, heat and safety rules limit what can run onboard.

Four things came together

• Edge GPUs such as Jetson Thor (kit launched August 2025).
• Vision-language-action models that turn images and words into motion.
• GPU simulation that trains thousands of robot copies at once.
• Synthetic data from world models to stretch scarce real data.

Hear and see

• Cameras and depth sensors stream images.
• The instruction arrives as text or speech.

Understand

• A VLA model on the onboard GPU reads images plus the words.
• About 10 to 30 times a second.

Plan

• Output is a short action sequence.
• Planners check reach and collisions.

Balance and move

• Whole-body control turns targets into joint commands.
• Tens to hundreds of times a second.

Drive motors

• Servo drives close torque and position loops.
• Thousands of times a second over EtherCAT.

Learn

• Failures and rescues are logged.
• Fine-tuned weights return overnight.

Slow reasoning sits on top and fast reflexes at the bottom. A robot is three programs, not one.

The learning loop closes the circuit: what the robot gets wrong today is training data for tomorrow.

degrees of freedom in the 2026 production Atlas

humanoid pilots at GXO, not yet paid back (Aug 2026)

Built for human spaces

• Doors, stairs, shelves and tools are made for people.
• No need to rebuild the site around the robot.
• One body for many tasks.

Learning from people

• Human video and teleoperation map more directly to a human-shaped body.
• Demonstrations by workers become training data.
• Operators can steer the robot intuitively.

What it costs

• Balance on two legs adds control, power and fall risk.
• Many joints mean more parts that can fail.
• Safety certification for dynamically stable robots is still being written.
• Wheels on a torso often do a task more cheaply.

Motor plus gearing

• Brushless motors with harmonic drives, planetary or cycloidal gearboxes.
• Quasi-direct drives keep gearing low for good torque feel and shock tolerance.
• Encoders and torque sensors give the controller its sense of position and load.

Strain-wave (harmonic) gear set

Photo: Pieceofmetalwork, CC BY-SA 4.0

Screws and rods

• Roller-screw linear actuators push and pull for legs and arms.
• Strong and compact at the cost of added mass in the limb.
• Common in several humanoid designs.

2026 production Atlas

• 56 degrees of freedom, most joints rotate 360 degrees.
• The design removes fragile wiring from moving joints.
• Hyundai Mobis supplies the actuators.
• Atlas moved from hydraulic to electric drive.

Many joints in a small space

• A human hand has about 27 degrees of freedom.
• Robot hands with 20 or more actuated joints pack motors, tendons and wiring into a palm.
• More joints mean more cost, heat and wear.

Grippers do most paid work

• Two-finger and suction grippers handle most logistics picking today.
• Three-finger and dexterous hands are for tool use and fine assembly.
• Tool changers swap end effectors when a task needs one.

Touch closes the loop

• Vision cannot see grip force or slip.
• Tactile skins and camera-based gel sensors report contact.
• Touch data is scarce, which limits what models learn about it.

Runtime and charging

• Lithium-ion packs are the norm. Runtime of a few hours per charge is typical for humanoids.
• Swappable packs or automatic docking turn a demo into a shift.
• Battery safety needs its own review.

Motors get hot

• Peak torque is short. Continuous torque is set by how well the motor sheds heat.
• Hot joints derate, slow down or shut off.
• Enclosures and cooling add weight and noise.

Weight changes everything

• Doubling payload (say 10 kg to 20 kg) changes motors, gearing and structure.
• Simulation can test feasibility.
• It cannot test sourcing, safety certification or reliability.

Design in CAD

Autodesk, PTC Onshape or SolidWorks define parts, joints and mass.

Export

STEP for geometry, URDF or MJCF for joints, USD for full scenes.

Simulate

Load into Isaac Sim or MuJoCo. Test reach, balance and payload.

Prototype

3D print light parts. Buy actuators, drives, sensors and batteries.

Certify

Safety testing and standards before any customer site.

Pilot

Staged live trial. Rescues feed the next design.

The chain echoes the CAD-to-CNC shift of the 1990s, with a simulation step added in the middle.

Build to order is real for prototypes and light parts. Production humanoids still depend on bought-in actuators and sensors.

Runs the AI models

• Edge GPUs such as NVIDIA Jetson Thor run vision and VLA models on the robot itself.
• The Thor developer kit launched in August 2025 at $3,499; NVIDIA raised its list price to $5,499 in July 2026 as memory costs rose.
• Smaller Thor and Orin modules suit lighter robots.

Runs the software

• ROS 2 nodes, planning, state estimation and logging.
• Real-time Linux or a real-time core keeps timing predictable.
• Isaac ROS packages are tuned for NVIDIA hardware.

Runs the motors

• Microcontrollers and drives close the fast loops.
• They must answer on time every time, so they stay simple.
• Safety functions often live on their own certified controller.

• NVIDIA Jetson AGX Orin developer kit (right) and its box (left).
• Orin is the generation before Thor; both put a GPU and CPU in a module small enough for a robot's torso.
• The same module, without the case, is built into production robots.
• Thor is the 2025 successor, aimed at humanoids and VLA models.

Capture

Cameras, depth, LiDAR, IMU and joints stream data.

Calibrate and sync

• Fix sensor positions and align timestamps.
• A one-frame error breaks grasps.

Perceive

Detect, segment and estimate pose of objects and people.

Localize and map

Where am I, what is around me. SLAM or a prebuilt site map.

World state

A live model of objects, free space and risks handed to planning.

ROS 2

• Open-source messaging framework that links sensors, planning and drivers. Middleware, not an operating system.
• Began at Stanford and Willow Garage in 2007; now stewarded by the Open Source Robotics Foundation.
• ROS 2 left beta in December 2017. ROS 1 reached end of life in May 2025.

EtherCAT

• Open real-time standard from Beckhoff, managed by the EtherCAT Technology Group.
• Servo drives stay synchronized to under a microsecond.
• Very common in industrial arms and humanoid joints.

Isaac ROS and others

• NVIDIA ships Isaac ROS, ROS 2 packages tuned for its hardware.
• The GPU does not drive motors directly. Commands pass through ROS 2 and the drive network.
• Alternatives such as CAN buses appear in smaller robots.

~10–30 Hz

AI brain

A VLA model reads camera images and an instruction such as "restock the soup" and outputs short action sequences. NVIDIA reports GR00T N1.7 at about 25 Hz on Jetson AGX Thor with optimized inference.

Edge GPU

10 to 500 Hz

Whole-body control and planning

Balance, walking, arm paths and obstacle avoidance. Often a reinforcement-learning policy trained in Isaac Lab, or model-predictive control, running as ROS 2 nodes.

CPU or GPU, ROS 2

1 to 10 kHz

Motor loops

Servo drives close the torque, speed and position loops. EtherCAT links them, synchronized to under a microsecond.

Drives and real-time controller

Optimization-based control

• Model-predictive control and whole-body optimization plan balance and contact forces.
• Explainable, predictable, tunable by engineers.
• Needs accurate models and compute.

RL locomotion policies

• Trained in Isaac Lab with thousands of parallel copies and randomized conditions.
• Handle rough ground and pushes that are hard to hand-code.
• Quality depends on how well simulation matches reality.

Arm paths and obstacles

• MoveIt 2 and GPU planners such as cuMotion compute collision-free arm paths.
• Planning runs below the AI brain and above the motors.
• Real sites add moving people and clutter.

Stops and limits

• Emergency stops cut power regardless of what the model wants.
• Joint, speed and force limits are enforced in firmware.
• Safety-rated controllers are separate from AI compute.

Collaborative modes

• Speed and separation monitoring slows or stops when a person is near.
• Power and force limiting caps contact force.
• These ideas come from industrial collaborative-robot rules.

A collaborative arm palletizing without a fence

Photo: CollaborativePalletizer, CC BY-SA 4.0

Functional safety

• Safety functions carry performance-level or SIL ratings.
• A learned policy cannot be certified directly, so the guard around it is.
• Legged and humanoid robots lack settled standards today.

Inputs

Camera frames, joint state and a text instruction.

Backbone

A pretrained model understands scene and words.

Action head

Generates a chunk of joint targets, often by diffusion or flow matching.

Action chunk

A short sequence executed while the next one is computed.

Controllers

Lower loops turn targets into torques.

NVIDIA GR00T N1.7 has open weights and about 3 billion parameters, and runs at about 25 Hz on Jetson AGX Thor.

Demonstrations recorded on real robots are the data used to fine-tune a VLA for a new task.

Start from web-scale knowledge

• A vision-language model already understands objects, words and scenes.
• Robot data then teaches it to act.
• This is why a few demonstrations can go a long way.

Demonstrations and trial

• Imitation from human demonstrations gives a first policy.
• Reinforcement learning in simulation refines it.
• Staged live operation adds real-world corrections.

Make it yours

• NVIDIA suggests about 100 to 500 good demonstrations for a fine manipulation task.
• Cross-embodiment models adapt across robot bodies.
• A body change may force retraining.

NVIDIA Cosmos

• A world foundation model that makes realistic video-like scenes.
• Cosmos Transfer adds variation such as lighting and clutter to simulated scenes.
• Cosmos 3 (June 2026) unifies reasoning, world and action generation.

JEPA-style models

• Yann LeCun argues for predicting in abstract latent space, not pixels, because most pixel detail is unpredictable.
• His AMI Labs raised a $1.03 billion seed round in March 2026.

What world models do for robots

• Create synthetic training data from scarce real data.
• Test policies safely before deployment.
• Possibly plan by imagining outcomes.
• Store environments in Isaac Sim as 3D OpenUSD scenes.

Agent chooses skills

• An LLM agent turns "tidy the aisle" into a sequence of existing skills.
• AWS's open-source Strands Agents and Strands Robots let an agent call skills, including over ROS 2.
• The agent plans; tested skills do the moving.

A library of tested behaviors

• Pick, place, open, navigate, scan.
• Each skill is a trained policy or a scripted routine with known limits.
• The library grows with every deployment.

Keep the agent in bounds

• Permissions limit which skills can run where.
• Human approval for risky steps.
• Logs show why the robot acted.

• NASA's Valkyrie humanoid fixing a solar panel in a Gazebo scene built for NASA's Space Robotics Challenge.
• Teams wrote and tested their control software against this virtual robot before any hardware run.
• The same idea scales up in Isaac Sim and Isaac Lab, with thousands of copies running at once.

Describe the task

Points for stocking a shelf correctly. Penalties for dropping glass.

Load the robot

Import the model into Isaac Sim. Add the site as a 3D scene.

Run thousands of copies

Isaac Lab trains in parallel on GPUs.

Randomize

Vary lighting, friction, mass and sensor noise.

Policy

A network that maps sensor input to actions.

Test on hardware

Deploy to the robot. Measure the gap.

A policy is the learned network. Rules like "do not drop glass" are only the rewards that shaped it.

Domain randomization is the core defense: train across many versions of reality so the real one looks like one more variant.

Where reality differs

• Gravity and rigid-body motion are well understood.
• Friction, soft or slippery objects, sensor noise and clutter are approximated.
• Unscripted events such as spills and dropped cans are hard to script.

How teams close it

• Domain randomization and system identification of real motors.
• Teacher-student distillation and residual policies.
• Physics-informed neural networks for better friction models (active research).

Keep real data in the loop

• Fine-tune on a few hours of real demonstrations.
• Staged live operation reveals what simulation missed.
• Full sim-to-real is still an open research problem.

A virtual site

• A physics-accurate virtual copy of a robot or site.
• Test layouts, routes and failures before touching the real floor.
• Built as 3D OpenUSD scenes in Isaac Sim.

Variation on demand

• Cosmos Transfer adds lighting, weather and clutter variants.
• One real clip becomes many training scenes.
• Cosmos augments data. It does not replace images of the real site.

What synthetic cannot do

• It cannot invent physics the simulator lacks.
• Quality depends on assets, called SimReady assets, that behave like real ones.
• Real data still anchors the final policy.

Real robot data

Real robot data on your task

Rescues and autonomous runs on the actual site. Scarce, costly, decisive.

Teleoperated demonstrations

Teleoperated demonstrations

Human-driven runs via headset, leader arms or gloves. Often 100 to 500 episodes for a new task.

Human video

Human video

Workers filmed doing the job, often from head-mounted cameras. Cheap, but it needs mapping to the robot body.

Simulation and synthetic data

Simulation and synthetic data

Isaac Lab runs and Cosmos-generated scenes. Unlimited volume, imperfect physics.

Internet video, images and text

Internet video, images and text

Pretraining of the vision-language backbone. Widest, least robot-specific.

How demonstrations are made

• XR headsets: NVIDIA Isaac Teleop supports Meta Quest 3, PICO 4 Ultra and Apple Vision Pro, with CloudXR streaming.
• Leader-follower arms, gloves and handheld grippers.
• NVIDIA supplies the software; the headsets are standard consumer devices.

How much is enough

• NVIDIA GR00T guidance: about 100 to 500 good episodes for a fine task.
• Harder tasks and new bodies need far more.
• Quality matters: failed or sloppy demos teach failure.

Teleoperation as a safety net

• A remote operator steps in when the robot stalls.
• Each rescue is recorded and becomes training data.
• Operators per robot is a key cost figure and should fall over time.

From shift to training set

• Record sensor streams, actions and outcomes.
• Keep failures and stuck moments. Not every second of video needs to flow back.
• Store in object storage such as S3, version, label and curate.

Fine-tune and redeploy

• Fine-tune on SageMaker or similar.
• Push improved weights to the robot overnight.
• Keep the raw data: a redesign can be retrained from it rather than from zero.

What to measure

• Task success rate and share finished without a human rescue.
• Interventions per hour and mean time between failures.
• Cost per robot-hour and payback against a human worker.

Survey and twin

Map the site. Build a digital twin. Pick one scope, one site, one metric.

Stage off-hours

Run in a bounded area while a remote operator steps in on failures.

Supervised shifts

Staff nearby. Track rescue rate and cycle time.

Add daytime hours

Expand only as confidence is earned.

Autonomous operation

Exceptions go to remote operators. Fleet monitoring on.

Improve

Fine-tune, update, repeat.

Confidence is earned shift by shift. Skipping staging is the most common reason a pilot fails.

Each rescue is recorded and used to fine-tune the model. The robot you stage is not the robot you run six months later.

Many robots, many vendors

• Open-RMF coordinates traffic and tasks across robots from different makers.
• OpenRobOps (InOrbit, released September 2026, Apache 2.0) is an open fleet manager with the first reference implementation of ISO 21423.
• Vendor consoles still dominate.

Remote operators

• An operator answers help requests from a queue.
• The ratio of operators to robots drives cost.
• Staffing and shift planning matter as much as software.

Updates and uptime

• Over-the-air updates for software and models, with rollback.
• Spare parts, battery swaps and repair turnaround.
• Monitoring flags drift before a failure.

A fleet of driverless pallet stackers

Photo: Carmenter, CC BY-SA 4.0

Software you must patch

• The EU Cyber Resilience Act phases in obligations from 2026, with main duties in 2027.
• Plan for secure boot, signed updates and vulnerability reporting.
• A hacked robot is a physical risk.

Cameras on the floor

• Robots record workers and customers.
• Consent, retention and masking rules vary by region.
• Opt-in data sharing needs clear contract language.

Insurance and responsibility

• Who is liable when a robot causes harm: maker, integrator or operator.
• Insurers ask for safety evidence and incident logs.
• The insurance position is a diligence question.

1

Mechanics

• The 2026 Atlas has 56 degrees of freedom, 360-degree joints and no fragile wiring in moving joints.
• Its 2026 run is committed to Hyundai and Google DeepMind.

2

Data and the model

Tesla's fleet and LG's data factory are bets that proprietary real-world data is the moat.

3

Customer integration

Vertical ties shorten the path to revenue: Boston Dynamics with Hyundai, Agility with Toyota and GXO.

4

Service, safety, uptime

• Often the least glamorous and hardest part.
• It decides whether a pilot becomes production.

GPUs and Jetson

• Access to GPUs for training.
• Jetson chip runs the result on the robot.

CUDA, Omniverse, OpenUSD

• Real-time graphics and physics heritage.
• Isaac Sim, Isaac Lab, Isaac ROS, GR00T and Cosmos sit on it.

Partners plug in

• PTC, Cadence, Siemens and AWS connect to the stack rather than rebuild it.
• Newton physics is shared with Google DeepMind and Disney Research.

Dexterous hands

Reliable, affordable hands with touch.

Reliability on the long tail

Rare events dominate real shifts.

Contact-rich sim-to-real

Friction, soft and slippery objects.

Data at scale

Real demonstrations are costly and scarce.

Safety certification

Standards for legged and humanoid robots.

Unit economics

GXO: 45 humanoid pilots, none paid back yet.

Operator ratios

Fewer humans per robot over time.

Energy and uptime

Runtime, charging and heat.

What is not the standard stack?

Actuators, hands, model or data that the company builds itself.

Rescue rate today

Share of tasks finished without a human rescue.

Real versus simulated hours

Demonstration hours held in each.

Who owns data and weights?

Can it run in the customer's own cloud account?

Operators per robot

Is the ratio falling?

Cost per robot-hour

All-in, and payback against a human worker.

Last pilot failure

Was it a sim-to-real problem?

Redesign impact

What must be retrained if hardware changes?

Safety and insurance

Certification and insurance position.

Consulting engagement

• Map the stack for the client and test it with the nine diligence questions.
• Check site readiness: power, network, safety, people.
• Review the maker, the model supplier and data rights.
• Pick one scope, one site, one metric.

Short pilot

• Usually 4 to 8 weeks: off-hours staging, then supervised shifts.
• Track rescue rate, cycle time and cost per robot-hour.
• A small senior pod runs the site, with a named human manager for every robot.
• No capital outlay for machines.

Production

• Managed crew, outcome-based or advisory terms.
• Data loop and fleet operations in place.
• Remote operators and spare parts on call.
• Scale up as results are proven.

Safety and law

• EU Machinery Regulation applies 20 January 2027.
• ISO 25785-1 progress for legged and humanoid robots.
• Cyber Resilience Act obligations phase in.

Models and data

• Newer VLA and world models, open and closed.
• Cheaper edge compute.
• Progress on hands and touch.
• Larger consented data sets.

Proof of ROI

• Do pilots convert to production at GXO and others?
• Does RaaS pricing hold?
• Do operator ratios fall?
• Who owns the data?

A robot is a stack

Seven layers from body to fleet operations. Few companies own more than two of them.

Three loops, three speeds

AI brain at about 10 to 30 Hz, whole-body control at hundreds, motor loops at thousands of Hz.

The plumbing is shared

ROS 2, EtherCAT and NVIDIA's Isaac tools are common. Makers differ on mechanics, data and model.

Hands and power limit the shift

Dexterous hands, battery runtime and heat decide real work more than any demo.

VLA models are the new brain

Images and words in, actions out. Open (GR00T, π0) and closed (Gemini Robotics, Helix) families.

Simulation trains, the site finishes

Domain randomization narrows the gap. Contact-rich tasks still need real data and staging.

Real data is the moat

Teleoperated demonstrations and rescues on the target robot are scarce and decisive.

Data rights shape the contract

A base model plus a thin local fine-tune, run in the customer's own cloud where needed.

Safety sits below the AI

Certified stops and limits guard the model. Humanoid standards such as ISO 25785-1 are still in draft.

The business is service

Uptime, rescue rate, operators per robot and cost per robot-hour decide returns.

What Atomgigs is

• Robots, drones, VTOL, IoT and automation, with people, as one hybrid workforce.
• Humans in the lead, physical AI in the loop.
• Not a machine builder: it sources from makers, then staffs, integrates, deploys and operates.

The sources

• Robots: every robot type, profiled in depth.
• Everything other than robots: drones, VTOL, IoT and automation, from many makers.
• The humans: role families across the trades, engineering and safety, every person a contractor.

Who we work with

• Investors: consulting before and after the deal.
• Enablers: financiers, insurers, platforms and regulators behind every deal.
• Channel partners: engineering, design, MEP and construction firms.

Buyer groups

• Data centers
• Manufacturing and logistics
• Construction and infrastructure
• Public safety and defense (no weapons)
• Commercial services
• Physical AI labs

How it works

• Consulting engagement, then a short pilot, then production.
• One scope, one site, one metric in the pilot.
• A small senior pod runs every site, with a named human manager for every robot.

How we are paid

• Pilot, managed crew, outcome-based or advisory.
• No capital outlay for machines and no long-term headcount.
• Pay for capacity and outcomes; scale up as results are proven.

Mavengigs Consulting

• Strategic M&A, transformation and optimization
• Focused on the part of M&A that begins after everyone shakes hands
• Built for PE-backed and Fortune 500 teams

Mavengigs Capital

• An independent sponsor platform
• Family office strategy and management

Mavengigs Ventures

• Funding and mentorship for founders we back
• Panmedical: digital health in India, run as CEO
• Atomgigs, Mavens Life and Mavens Health

Career and experience

• 30+ years of management consulting: managed major M&A deals totaling $20 billion; scaled 50+ companies for $100 million of impact
• Deep expertise in M&A consulting, including integrations and divestitures
• 2018 to present: Mavengigs: Consulting, Capital and Ventures
• 2004 to present: Founded Panvisage, CEO for 22 years
• 2024 to present: Panmedical CEO (digital health)
• 1998 to 2004: Oracle, led 14 enterprise programs
• 1994 to 1996: Rolled out nationwide cell networks
• 1992 to 1994: Car and truck design (CAD, CAM, CIM)
• 1991: B.E. Electrical and Electronics Engineering
• 1989 to 1991: Designed robots, AGVs and manufacturing transfer lines
• Data centers: built one from scratch in 2000; data center migration in 2008; Equinix in 2020
• E&C companies: WSP, Jacobs, Worley, Halliburton, Fluor, Brown & Root
• Managed the buildout of a manufacturing plant and a distribution center

Monty Pandey

• Founder and CEO, Atomgigs
• Founder and CEO, Mavengigs
• Co-founder and CEO, Panmedical

Not investment advice

Atomgigs provides consulting. We are not an investment advisor, so you make your own decisions on investments.

Products belong to their owners

Company, product and robot names, logos and photos are used as examples and belong to their owners. This deck is for information only and is not a contract or an offer.

About the numbers

Numbers come from publicly available sources and are the latest available at publication. Costs, prices and returns are typical ballpark values, not quotes; they will be updated for your specific engagement.

• NVIDIA newsroom, blogs and developer docs: Jetson Thor, Isaac Sim, Isaac Lab, Isaac ROS, Isaac Teleop, GR00T N1.7, Cosmos 3 (June 2026), Newton
• NVIDIA Jetson price changes, July 2026 (NVIDIA developer FAQ, as reported by CNX Software)
• AWS: Physical AI Toolchain (Oct 2026), Amazon Bedrock data privacy, Strands Agents and Strands Robots
• Boston Dynamics and Hyundai: production Atlas (CES 2026)
• Physical Intelligence (openpi), Google DeepMind (Gemini Robotics 2), Figure (Helix), OpenVLA
• Open Robotics: ROS 2, Gazebo and Open-RMF documentation; Google DeepMind: MuJoCo

• GXO Logistics, Q2 2026 earnings call (Aug 2026); Agility Robotics and Toyota Motor Manufacturing Canada (Feb 2026); Agility merger filing (2026)
• InOrbit: OpenRobOps release (Sept 2026)
• ISO 10218-1 and -2:2025, ISO/CD 25785-1, ISO 13482, ISO 3691-4, ISO 21423; Regulation (EU) 2023/1230; EU AI Act and its 2026 amendment; EU Cyber Resilience Act
• Omdia humanoid shipments, 2025 (Jan 2026)
• AMI Labs funding (Mar 2026); NVIDIA and LG Group AI factory (June 2026)
• Robots-as-a-Service pricing: vendor pages (Brightpick, 1X) and industry benchmarks