Executive briefing

Week of 26 October 2026

AI is taking on physical work. Every action has a time budget.

Physical AI uses what a machine senses to guide how it moves and acts.

Jensen meets the Microducks.

Microducks are small robots from Pollen Robotics. Developers train their AI to walk and balance in a simulated world, then run it on the real robot to control its movements. Here, NVIDIA CEO Jensen Huang meets them at GTC Berlin.

NVIDIA official Microducks demonstration at GTC Berlin 2026

The robot adjusts its movements to stay balanced.

Microduck by Pollen Robotics

Watch a robot work through a kitchen.

Figure’s humanoid robot uses an AI system called Helix 02 to coordinate walking, balancing and handling objects. Camera and touch feedback guide its movements as it unloads and reloads a dishwasher.

Figure Helix 02 kitchen demonstration, three original-speed excerpts

From dishwasher to cupboard: carry, reach and place.

Physical AI can combine movement and manipulation into a longer job.

Figure’s kitchen demonstration

Watch the task. Compare the time and cost.

Artificial Analysis takes a different approach: it gives general-purpose AI models camera views and code to control a robot arm. Watch OpenAI’s Astra model attempt three tabletop tasks, then compare the result, elapsed time and model cost.

GPT-6 ASTRA · AA ORIGINAL RECORDING

Pour into the cup without spilling.

Task completed

7 min 30 sec

Recorded task duration

US$16.36

Average model cost per task

The drink reached the cup without a spill.

Model cost is the average across eight tasks. The recordings compress the elapsed time; 2× speeds them up further.

Inspect original run

The arm moves for only part of the job.

Understanding the scene, planning movements and checking results can take longer than moving the arm.

20%Understand
59%Plan movements
13%Move the arm
8%Review + other

Astra’s time across eight attempts, rounded. Reviewing images accounts for 6%; other activity, 2%.

Judge the result, the full attempt time and the model cost together.

The guest pulls away. The robot must keep holding.

A guest pulls their hand away while an AI model in a data centre plans the handoff. The robot must keep holding the cup. Compare waiting for the answer with keeping sensing and grip control running on the robot.

The guest reaches for the drink.

Remote AI · planning the next step

Review the handoff.

Upload → thinking → return
On the robot · sensing and grip

The guest has not taken the cup. Keep holding.

Send the image while the robot keeps its grip.

Check the new plan against the scene before the robot acts.

Upgrade the technology. Watch the service expand.

Reaching a guest, setting out a meal and handling a blocked table depend on different capabilities. Explore the possible hotel or MTR service to see what each capability adds, and where the job stops without it.

From doorstep delivery to personal room service.

The guest has an injured hand. Bring the meal within reach, adapt to the room, and collect only the dishes they specify.

0 / 5 capabilities enabled

What could this robot do next?

Delivery reaches the door.

The complete service needs these capabilities working together.

How it works

Navigation, handling, balance, planning and interaction solve different parts of the job.

A useful service needs these capabilities to work together, from reaching the task to confirming it is complete.

Keep the robot responsive while the larger model plans.

Use a larger model to plan the job while onboard systems keep checking position, grip and balance.

Remote data centre

Larger AI model · data-centre servers

Work out what the guest needs, compare options and plan the next service step.

Upload → think → return

Check the proposed action against the scene now

On the robot

Local AI chip · motor controller

Local AI processes live camera input. The motor controller uses sensor feedback to keep the grip secure and the robot balanced while the wider plan is prepared.

Sense → update → control

Plan the wider job while the robot stays responsive.

One question to take away

The next robotics proposal should show you a complete job.

What would a robot need to complete one useful job in your business, at the pace the work demands?