AGI Index dimension · weight 7%
Embodiment
Acting competently in unstructured physical space.
Perception and manipulation in the physical world outside controlled environments. The lowest-scoring dimension in the Index, and the one whose gap to the others is widening.
The argument
Evidence and counter-evidence
Both sides are published at equal weight. A framework that only records what raises a score is not measuring anything.
Raises the score
- General-purpose manipulation policies transfer across robot morphologies, which was not true recently.
- Vision-language-action models let robots follow open-ended instructions in semi-structured settings.
Holds it down
- Real-world reliability remains far below what deployment requires. Demonstrations are selected; failure rates in the field are not published.
- Data scarcity is structural — physical interaction data cannot be scraped, and collecting it is bounded by hardware and time.
- No score movement this quarter. The gap between digital and physical capability is widening, not closing.
What the score reads from
Unstructured manipulation
Cross-embodiment transfer
Embodiment over time
| Date | Index |
|---|---|
| Jan 2020 | 22.0 |
| Jul 2020 | 23.0 |
| Jan 2021 | 24.0 |
| Jul 2021 | 25.0 |
| Jan 2022 | 26.0 |
| Jul 2022 | 27.0 |
| Jan 2023 | 29.0 |
| Jul 2023 | 30.0 |
| Jan 2024 | 32.0 |
| Jul 2024 | 34.0 |
| Jan 2025 | 36.0 |
| Jul 2025 | 39.0 |
| Jan 2026 | 40.0 |
| Apr 2026 | 41.0 |
| Jul 2026 | 41.0 |
| Sep 2026 | 41.0 |
Related coverage
Everything bearing on embodiment
News
Research
Context
Where this sits against the rest
- Multimodality82
- Coding81
- Tool use76
- Reasoning74
- Memory67
- Generalization66
- Planning61
- Learning59
- Autonomy53
- Social intelligence52
- Scientific discovery48
How these scores are producedMethodology v1.2