---
pretty_name: RLE-Bench task05
tags:
- robotics
- libero
- robotwin
---

# RLE-Bench task05

Demonstration shards mounted by the task05 (nanoVLA) cells: 128x128 frames stored as flat
little-endian arrays with JSON sidecars. `rlebench prepare task05` fetches them at a pinned
revision and checks every file against `SHA256SUMS`.

| directory | source | tasks | episodes | frames | state / action dim |
|---|---|---|---|---|---|
| `shards_l10_128/` | LIBERO-10 from `physical-intelligence/libero` at `a4336d589d589045d1c56423ffdf3b88a0e19b1f` | 10 | 379 | 101469 | 8 / 7 |
| `shards_rt15_128/` | RoboTwin 2.0 aloha-agilex clean demonstrations (`clean50`) | 15 | 700 | 135519 | 16 / 14 |

In each directory, `agent.u8` and `wrist.u8` hold `frames x 128 x 128 x 3` uint8 images (for
RoboTwin, the two cameras `meta.json` lists), `state.f32` and `actions.f32` the per-frame vectors,
`ep_end.i32` the episode boundaries, `task_idx.u8` the task of each frame and `meta.json` the
shapes. The remaining files hold task strings and tokens, instructions, episode ids and
normalisation statistics.

The upstream datasets' licenses apply to the derived frames.