---
pretty_name: ERVLA AgiBot
language:
- en
tags:
- robotics
- embodied-chain-of-thought
- agibot
---

# ERVLA AgiBot

AgiBot demonstrations with embodied chain-of-thought annotations for
[ERVLA](https://arxiv.org/abs/2606.03784).
The demonstrations originate from [AgiBot World](https://agibot-world.com/).

| Content | Files |
| --- | ---: |
| Episode JSON | 776,158 |
| CoT JSON | 776,158 |
| Videos | 2,327,916 |

The dataset is distributed as independent `tar.zst` archives in `archives/`.
All archives extract into the same dataset root:

```text
json/
cot/
videos/
```

Paths in JSON and CoT are relative to this root. Archives contain the video
contents, with no symlinks or machine-specific ownership metadata.
Download every archive to restore the full dataset; individual archives may
contain only one content type. `archives.json` lists available archives and
their checksums. Its `complete` field becomes `true` after all archives have
been uploaded.

```bash
hf download ERVLA/AgiBot --repo-type dataset --local-dir agibot
cd agibot
sha256sum -c SHA256SUMS
for archive in archives/*.tar.zst; do
  tar --zstd -xf "$archive"
done
```

Extraction requires GNU tar and Zstandard (`zstd`).

End-effector projections were refreshed where camera calibration was
available; 2,145 CoT files retain their original projections.

[Project](https://taoshuaiz.github.io/ERVLA/) · [Paper](https://arxiv.org/abs/2606.03784)