Basic skills
Can it carry out a simple instruction?
90%
85–99% (sim.)
record in simulationline: product reliability (99%)
Record trend since Jun 2023
Grasping, placing, opening, putting away on request, in a familiar environment.
Intelligent robotics observatory
How far are we from a truly autonomous robot?
- record success rate in simulation on simple instructions
- 90%
- tasks completed by the best humanoid tested in 30 unfamiliar homes (Figure 03, Sep. 2026) source
- 56%
- time taken by an autonomous robot compared with a human on everyday tasks (Jan. 2026) source
- 4× to 10×
- long household missions completed end to end, in simulation (BEHAVIOR Challenge 2025) source
- 12.4%
- best public-evidence score among 71 humanoids (Figure AI Figure 03)
- 10 / 18
- benchmarks measuring human-level general autonomy
- 0
Computed from 3,023 published results for 1,320 models. Percentages are success ratesSuccess rateOut of 100 attempts, how many times the robot completed the task. on benchmarksBenchmarkA standardized test: the same tasks for everyone, so that models can be compared fairly. in simulationSimulationA virtual world (like a realistic video game) where tests can be run thousands of times without breaking real hardware..
Distance to an autonomous robot
1/ 5
level reliably reached, in the real world
- 0
- 1
- 2
- 3
- 4
- 5
programmed robot“Terminator”
Level 2 (follows instructions): 85–99% (sim.) · average 90% · threshold 99% and real-world evidence.
The autonomy ladder
From the programmed factory robot to the science-fiction robot. A level counts as reached only with two kinds of evidence: product-grade reliability (at least 99% success, the vertical line) on all of its capabilities, and a demonstration outside simulation, with real customers.
- Level 0Achieved
Programmed
Repeats a pre-coded movement, without perceiving or understanding anything.
Welding robot on a car assembly line.
- We are hereLevel 1Achieved
Learns a skill
Reproduces a movement demonstrated by a human, in a specific place.
Folding a towel after about a hundred demonstrations.
- Level 285–99% (sim.)
Follows instructions
Understands a written instruction and performs varied tasks in a familiar environment.
“Put the mug in the top drawer.”
- Level 350–85% (sim.)
Adapts
Remains effective when the scenery, the objects or the robot change; uses both hands; manages in an unfamiliar home.
Clearing the table in a kitchen it has never seen.
- Level 4< 50% (sim.)
Autonomous over time
Remembers, plans long missions and corrects its errors without human help.
Tidying the whole house over an afternoon, unsupervised.
- Level 5Not measurable
General autonomy
The science-fiction robot (the Terminator, Asimov's androids): any task, anywhere, for days, improvising in the face of the unknown.
No existing benchmark measures it.
No benchmark can measure it yet.
embodiedrank's reading grid, for guidance only: it is not an official standard. Records often come from different models, fine-tunedFine-tuningRetraining a model on the data of a specific test before taking it. for each benchmark — no robot yet holds all of these records at once.
Why 90% is not enough: the last 10 percent
A mission is a chain of steps, and each step can fail. Probability of finishing it without help = (success per step)number of steps. Between 90% and 100%, each point gained changes the outcome by an order of magnitude.
| Success per step | 10-step mission | 30-step mission | 100-step mission |
|---|---|---|---|
| 90% | 35% | 4.2% | < 0.01% |
| 95% | 60% | 21% | 0.59% |
| 99% | 90% | 74% | 37% |
| 99.9% | 99% | 97% | 90% |
Errors over a day of 200 actions
- 90%20 errors
- 95%10 errors
- 99%2 errors
- 99.9%0.2 errors (1 every 5 days)
Going from 90% to 99% cuts errors by a factor of 10; from 99% to 99.9%, by another 10. Each additional “9” represents the same factor of 10.
Simulator
90%
30
Chances of finishing the mission without help
4.2%
With a 90% success rate per step, over 30 steps, the robot finishes on its own 4 times out of 100.
Simplified calculation: independent steps and no error correction. A robot that can recover does better; a robot that causes damage does worse. The number of steps per mission is an order of magnitude.
Mission success rate by mission length
- Product target (99%)
- Simulation record (90%)
- Real, unfamiliar home (56%)
Show chart data: Chances of finishing the mission without help, by number of steps
| Number of steps | Product target (99%) | Simulation record (90%) | Real, unfamiliar home (56%) |
|---|---|---|---|
| 1 | 99% | 90% | 56% |
| 5 | 95% | 59% | 5.5% |
| 10 | 90% | 35% | 0.30% |
| 15 | 86% | 21% | 0.02% |
| 20 | 82% | 12% | < 0.01% |
| 25 | 78% | 7.2% | < 0.01% |
| 30 | 74% | 4.2% | < 0.01% |
| 35 | 70% | 2.5% | < 0.01% |
| 40 | 67% | 1.5% | < 0.01% |
| 45 | 64% | 0.87% | < 0.01% |
| 50 | 61% | 0.52% | < 0.01% |
| 55 | 58% | 0.30% | < 0.01% |
| 60 | 55% | 0.18% | < 0.01% |
| 65 | 52% | 0.11% | < 0.01% |
| 70 | 49% | 0.06% | < 0.01% |
| 75 | 47% | 0.04% | < 0.01% |
| 80 | 45% | 0.02% | < 0.01% |
| 85 | 43% | 0.01% | < 0.01% |
| 90 | 40% | < 0.01% | < 0.01% |
| 95 | 38% | < 0.01% | < 0.01% |
| 100 | 37% | < 0.01% | < 0.01% |
Human vs. best robot, point by point
Human reference values and the best robot result published in 2025–2026, with sources.
In simulation: the ideal robot vs. the records
100% on every axis = the ideal autonomous robot. Shaded area = best published results, in simulation only.
- Ideal autonomous robot
- Best published result
Show chart data: Scores by capability (0 to 100)
| Capability | Ideal autonomous robot | Best published result |
|---|---|---|
| Basic skills | 100% | 90% |
| Robustness | 100% | 81% |
| Realism | 100% | 93% |
| Two arms | 100% | 96% |
| Home | 100% | 74% |
| Memory | 100% | 52% |
| Long missions | 100% | 25% |
How to read this chart
Each spoke is a capability; 100% = perfect success. The shaded area = best result published in simulation for each capability. Records come from different models, often fine-tuned for each benchmark.
| Capability | Published record | Gap to 100% |
|---|---|---|
| Two arms | 96% | 4 pts to go |
| Realism | 93% | 7 pts to go |
| Basic skills | 90% | 10 pts to go |
| Robustness | 81% | 19 pts to go |
| Home | 74% | 26 pts to go |
| Memory | 52% | 48 pts to go |
| Long missions | 25% | 75 pts to go |
Capability by capability
What each benchmark checks, the current record and how it has progressed.
Robustness
Does it hold up when conditions change?
81%
50–85% (sim.)
record in simulationline: product reliability (99%)
Record trend since Oct 2025
Different lighting, camera, scenery or instruction: does it truly understand, or does it repeat by rote?
Realism
Would its results hold on a real robot?
93%
85–99% (sim.)
record in simulationline: product reliability (99%)
Record trend since May 2024
Simulations calibrated to predict performance on real machines.
Two arms
Can it coordinate two hands?
96%
85–99% (sim.)
record in simulationline: product reliability (99%)
Record trend since Jun 2025
Handing over, holding with one hand while acting with the other, manipulating with both arms.
Home
Can it manage in an unfamiliar home?
74%
50–85% (sim.)
record in simulationline: product reliability (99%)
Record trend since Jun 2024
Varied kitchens, everyday objects, common sense.
Memory
Does it remember what it has seen or done?
52%
50–85% (sim.)
record in simulationline: product reliability (99%)
Record trend since Feb 2025
Remembering a position, counting, resuming a task where it left off.
Long missions
Can it carry out a long mission on its own?
25%
< 50% (sim.)
record in simulationline: product reliability (99%)
Planning dozens of steps, handling the unexpected, replanning.