---
license: apache-2.0
---

# KODA Assets

Raw iPhone recordings and a fine-tuned ScaleBFM checkpoint for the [KODA pipeline](https://github.com/Gamma-Robot/KODA).
The original example remains available, and a second recording is now included.

| Recording | Download | File size |
| --- | --- | --- |
| `session_20260926_005030_iphone14_3` | [MCAP](https://huggingface.co/datasets/TeleEmbodied/koda-assets/resolve/main/samples/session_20260926_005030_iphone14_3/session_data_20260926_005030_iphone14_3.mcap?download=true) | 150,478,283 bytes |
| `session_20260926_004945_iphone14_3` | [MCAP](https://huggingface.co/datasets/TeleEmbodied/koda-assets/resolve/main/samples/session_20260926_004945_iphone14_3/session_data_20260926_004945_iphone14_3.mcap?download=true) | 127,547,978 bytes |

Each recording is distributed as its original MCAP file. The new `004945`
recording contains 367 RGB frames, 367 depth frames, camera poses and calibration,
IMU data, and mapping streams over approximately 12.29 seconds.

## Download the new example

With the Hugging Face CLI installed:

```bash
hf download TeleEmbodied/koda-assets \
  --repo-type dataset \
  --include "samples/session_20260926_004945_iphone14_3/*" \
  --local-dir data/koda-assets
```

The input file will be:

```text
data/koda-assets/samples/session_20260926_004945_iphone14_3/session_data_20260926_004945_iphone14_3.mcap
```

Follow the [KODA installation and pipeline instructions](https://github.com/Gamma-Robot/KODA#readme)
and select this MCAP as the input recording. File sizes and SHA-256 checksums are
listed in [`metadata.json`](metadata.json). Its existing `example` entry continues
to identify the original sample; the `examples` list describes both recordings.

## Fine-tuned ScaleBFM checkpoint

[`model_24600.pt`](https://huggingface.co/datasets/TeleEmbodied/koda-assets/resolve/main/models/scalebfm/model_24600.pt?download=true)
is the fine-tuned ScaleBFM checkpoint used for KODA's G1 torso/UMI3 tracking
experiments. See [ScaleBFM](https://huggingface.co/WeishuaiZeng/ScaleBFM) for the
upstream model.

| File | Size | SHA-256 |
| --- | --- | --- |
| `models/scalebfm/model_24600.pt` | 72,856,655 bytes | `2d40b7326c3435c9aaf01f46e1507f747e7a7371bfa504201aaea12c539f813d` |

From a KODA checkout with its dependencies installed, download and select it:

```bash
uv run hf download TeleEmbodied/koda-assets models/scalebfm/model_24600.pt \
  --repo-type dataset \
  --local-dir assets

export KODA_SCALEBFM_CHECKPOINT=assets/models/scalebfm/model_24600.pt
```

Then run `uv run koda run ...` in the same shell using the input and output paths
described in the [KODA README](https://github.com/Gamma-Robot/KODA#run-the-kinematics-and-video-pipeline).
The environment variable selects this checkpoint for whole-body tracking.

To check the downloaded file on Linux:

```bash
echo "2d40b7326c3435c9aaf01f46e1507f747e7a7371bfa504201aaea12c539f813d  assets/models/scalebfm/model_24600.pt" | sha256sum -c -
```