Lead story
Robot Hands for Modern AI and Real Work
Boston Dynamics · 2 Oct · Primary source · Boston Dynamics Atlas (electric)
Read the story
The latest
All news →
The Physical AI Power List
The people shaping what machines can do.
2026 Q3 quarter-to-date · Provisional
Rank 1
Elon Musk
Co-founder and CEO · Tesla
96Strategic Power /100
Joint first: unusually integrated humanoid resources, agenda influence and options balance NVIDIA’s broader enabling reach.
All 18 leaders
What changed · September 2026
The shape of the field.
3 moved up · 3 moved down · 7 held · 0 new
Capability landscape · September 2026
The whole field on one plot.
13 ranked
Select a robot
One evidence base · Three ways to use it
See the ranking
Every humanoid robot we track, ordered by what it can actually be shown to do.
Open the ranking Compare two robots →Check realness
Tell a demo from a deployment. Each score is linked to the public source behind it.
See the record How scores are built →Find fit
Start from the job you need done, then shortlist the machines that can realistically do it.
Find a robotThe current read · September 2026
A closer look at one robot.
Agibot (Zhiyuan Robotics)
Agibot A2
- Capability score
- 1.50
- Proof score
- 0.20
- Proof gap
- 1.30
Spider view · 8 axes
One spoke per axis. Dashed graphite is the Proof score; filled magenta is the Capability score. Gaps in either polygon mark axes that are not evidenced — never zeros.
0 of 8 axes not evidenced
Hover a spoke or axis label to preview it here. Click to open the axis lens.
31 of 40 sub-axes carry scored evidence in this edition.
Open the Agibot A2 reportHumanoids
What a humanoid actually does, step by step
- Step 01
Sees and senses
Cameras, microphones and touch sensors turn the room into one picture the robot can use.
- Step 02
Understands the task
It works out what it has been asked to do, and which objects in front of it matter.
- Step 03
Plans ahead
Before moving, it estimates what will happen next — where things fall, slip or get in the way.
- Step 04
Moves and handles
It chooses the actual movement: which hand, which grip, how much force.
- Step 05
Stays stable
Hundreds of times a second it adjusts the whole body so the work does not topple it.
- Step 06
Learns from experience
Between runs, recorded attempts are used to retrain the models so the next attempt is better.
Evidence
From vendor claim to real work
Six steps of proof
Each step is harder to fake than the one before it
Step 4 · Confirmed customer use
A real customer or site is named on the public record.
To move up: published output, uptime or task metrics.
No tracked record sits at this step with a public source we can link to yet, so no example is shown.
- Deployments tracked
- —
- Public records
- —
No sourced deployment records are available here yet.
Careers
Who is hiring to build this.
Four career paths
- Intelligence & Autonomy
- Motion & Hardware
- Systems, Validation & Deployment
- Product & Delivery