LJ

Lentin Joseph

CEO, RUNTIME Robotics · Author of 11 books on ROS and robotics

01

What Physical AI is, and how far it has come

Slides 3–8

02

The models powering real robots

Slides 9–14

03

Deployments, plus China and India

Slides 15–22

04

Business models, enablers and hard problems

Slides 23–26

05

Taking a model to production

Slides 27–36

06

Outlook, takeaways and how to start

Slides 37–40

Sense

Cameras, depth, lidar, force and touch sensors

Reason

VLA models, world models and fleet models

Act

Real-time commands for arms, legs, wheels and grippers

Improve

Teleoperation, simulation, synthetic video and fleet logs feed retraining

LLM

Large Language Model

In: text
Out: text

e.g. GPT, Claude, Llama

“Pick up the red cup” → writes the steps

+ vision

VLM

Vision-Language Model

In: images + text
Out: text

e.g. PaliGemma, Qwen-VL, LLaVA

“Pick up the red cup” → “It’s on the left of the table”

+ action

VLA

Vision-Language-Action

In: images + text + robot state
Out: robot actions

e.g. π0, OpenVLA, GR00T

“Pick up the red cup” → moves the arm and grasps it

Input

Observe

Camera images, instruction, robot state

See

Vision encoder

SigLIP, DINOv2

Think

Language backbone

PaliGemma, Llama, Qwen, Eagle

Act

Action head

Flow matching, diffusion or action tokens

Output

Action chunk

e.g. the next 50 joint targets

1 · In production

Warehouses, logistics and robotaxis

Amazon, Agility × GXO, Ultra, Dyna Robotics, Waymo

2 · Paid pilots

Humanoids in factories

Figure × BMW, Atlas × Hyundai, Apptronik, Hexagon AEON, Skild × Foxconn

3 · Announced

Robots in the home

1X NEO is on pre-order; no confirmed volume deliveries yet

You send

Camera images, robot state and a task

Dream API

Dream-1

You get

Actions for your robot to execute

Works across robots

Runs on its own bimanual YAM and popular low-cost arms; a few hundred demos adapt it to a new body.

Zero-shot claims

The company says it handles new tasks and new scenes it has never seen.

Built for developers

API access, a remote Robot Playground and fine-tuning that returns a model ID.

How it works

1. Type a task, like “Put the fork on the plate”
2. Pick the objects and the robot body, such as YAM
3. Dream-1 turns camera images and the prompt into actions
4. Top-voted prompts from viewers run next

Try it: dreamlabs.ai/playground

Fleet and manipulation

DeepFleet coordinates 1M+ robots; Vulcan picks items with a sense of touch.

Logistics humanoid

Multi-year robots-as-a-service deal; moving totes in Georgia, reportedly 100k+.

E-commerce packing

Live order packing running Physical Intelligence’s π0.6 model.

Service robots

Folding, food prep and packaging, about 16 hours a day at hotels and laundromats.

With NVIDIA and Accenture

Jetson-based autonomous forklifts for GXO, trained in warehouse digital twins.

Robotaxis

500k+ paid rides a week (Mar 2026), driven by the Waymo Foundation Model.

1M+

robots deployed across Amazon’s operations network

~10%

expected gain in fleet travel efficiency from DeepFleet

1st

Amazon robot with a sense of touch: Vulcan

Home humanoid

Pre-orders open with a $200 deposit. Deliveries were promised for 2026; none confirmed yet.

Low-cost humanoid

Starts at $13,500. Sold mostly to research and development labs today.

• Unstructured, ever-changing spaces
• People, children and pets nearby
• Price must fit a household budget

86–97%

of global humanoid shipments in H1 2026 came from Chinese makers

~20k

humanoids shipped worldwide in H1 2026, about 3× a year earlier

$9B

Unitree’s IPO valuation; listed on Shanghai’s STAR Market, Aug 2026

AgiBot

Unitree

Galbot

UBTech

Leju

Rest of world

Reliance-backed · Noida

Launched Elixis-W, a wheeled industrial humanoid, at LogiMAT India 2026; about 100 units planned for supervised pilots.

Bengaluru

Vision-first robots that pick unknown objects from cluttered bins, a long-standing hard problem.

Bengaluru

Autonomous vehicles for messy automotive factory floors and outdoor industrial yards.

Build the whole robot

Figure, Tesla, 1X, Dyna

Own the hardware, the model and the data loop. Tighter integration, but every robot is built in-house.

License the brain

Physical Intelligence, Skild AI, NVIDIA GR00T, Dream Labs

One model across many robot makers. Scales faster, but depends on partners’ hardware and data.

Actions must arrive within milliseconds, on the robot, every time.

A wrong action can damage equipment or injure people.

No internet-scale robot dataset exists; teleoperation and simulation fill the gap.

Skills learned in simulation must survive real friction, lighting and wear.

Build · stages 1–3

Prove · 4–5

Run · stages 6–8

• Scene envelope: objects, lighting, clutter and how much they vary
• Task metrics: success rate, cycle time, units per hour
• Reliability: mean time between interventions (MTBI)
• Safety envelope: speeds, forces, zones, people nearby
• Unit economics: cost per task vs. people or fixed automation

• Teleoperation: leader-follower arms (ALOHA, GELLO) or VR
• Simulation: Isaac Lab or MuJoCo with domain randomization
• Synthetic video: world models such as NVIDIA Cosmos
• Fleet logs: ROS 2 recordings (rosbag2, MCAP) and interventions
• Versioning: LeRobot or RLDS datasets, versioned like code

Step 1

Simulation

Proves: task logic across thousands of randomized scenes.

Exit: success target met on randomized episodes.

Step 2

Lab bench

Proves: real sensors and contact physics; the sim-to-real gap.

Exit: target met on a fixed real test set and held-out objects.

Step 3

Shadow mode

Proves: behavior on the live line. The model proposes; people or the legacy system act.

Exit: proposals agree with what was actually done.

Step 4

Supervised pilot

Proves: throughput and reliability in real production.

Exit: intervention rate and cycle time meet the spec.

L1

Foundation model

Proposes action chunks. Untrusted: may be wrong at any moment.

L2

Action guard

Clamps joint limits, speed and workspace; checks collisions before execution.

L3

Real-time controller

Tracks trajectories at roughly 500 Hz–1 kHz; rejects infeasible commands.

L4

Safety-rated functions

Speed and separation monitoring, power and force limiting, safe torque off. Certified.

L5

Hardware

E-stop, brakes, physical guarding. Certified.

Three loops, three rates

Slow reasoning, fast reflexes: the fast loops never wait on the slow ones.

Make it fit

• Quantize FP16 → FP8/INT8 with TensorRT; re-test success after each step
• Chunk actions: compute the next chunk while this one executes (π0: 50-step chunks)
• Distill big models into small student policies where deadlines are tight
• Stay on-robot for perception and control; use the cloud for logs and retraining
• Watchdog: no fresh action by the deadline means a safe stop

1 robot

1 work cell

1 site

Whole fleet

1 · Deploy the current version

2 · Log every episode: video, state, actions

3 · Flag interventions and failures

4 · Curate and label the hard cases

5 · Retrain or apply RL from experience

6 · Re-run the ladder, then back to 1

Task spec with measurable targets, signed off

Risk assessment done; safety functions certified

Versioned dataset that includes failures and edge cases

Model meets its control deadline on the robot

Regression suite: sim plus real held-out scenes

Watchdog and safe-stop fallback tested

Shadow-mode results meet the spec

Every episode and intervention is logged

Pilot hits its intervention-rate target

Staged OTA rollout with tested rollback

01

02

03

04

05

1

Physical AI is in production where spaces are structured and data is plentiful: warehouses and logistics.

2

Humanoids are in long paid pilots; the next year tests whether they become fleet orders.

3

The stack is splitting into robot makers, model licensors and a simulation and compute layer beneath both.

4

Production is a loop with safety gates: bound the model, measure interventions, and let fleet data improve it.

Student or maker

• Build a low-cost SO-101 arm and record demos with LeRobot
• Train ACT or SmolVLA on your own task
• Prompt real robots in the Dream Labs playground

Robotics engineer

• Fine-tune openpi (π0) or GR00T N on your own demos
• Evaluate in Isaac Lab or MuJoCo before hardware
• Wrap the policy as a ROS 2 node behind an action guard

Company

• Pick one structured, repetitive task with a clear metric
• Run it through the 8-stage loop in a time-boxed pilot
• Track MTBI and cost per task from day one

Read with care

Most figures come from industry trackers and vendor announcements, not audited filings. Unit counts and model versions differ between sources. Check primary announcements before relying on a specific number.

Checked October 2026

Waymo, Figure at BMW, ISO 25785-1, China and India. Shipment trackers disagree, so the China slide shows ranges, not single numbers.

Lentin Joseph

CEO, RUNTIME Robotics

Author of 11 books on ROS and robotics, including Mastering ROS 2 for Robotics Programming

github.com/runtimerobotics