---
library_name: lerobot
tags:
- act
- robotics
- lerobot

---

# act_block_stack

Standard LeRobot ACT trained from scratch on SO101 data.

| | |
|---|---|
| **Policy** | ACT (`policy.type=act`) |
| **Dataset** | [villekuosmanen/armnetbench_block_stack](https://huggingface.co/datasets/villekuosmanen/armnetbench_block_stack) |
| **Task** | `block_stack` |
| **Action dim** | 6 (single-arm) |
| **Cameras** | `top`, `wrist`, `front` |
| **Training** | GCloud A100-80GB (alpha-act-a100-1), batch 16, action horizon 30 |
| **Loss @ 10k** | 0.163 |
| **W&B project** | [act_block_stack](https://wandb.ai/pravsels/act_block_stack) |
| **W&B run** | [6hzd0rjl](https://wandb.ai/pravsels/act_block_stack/runs/6hzd0rjl) |

## Checkpoints

Archived copies:

| Step | Path |
|------|------|
| 010000 | `checkpoints/010000/pretrained_model/` |
| 008000 | `checkpoints/008000/pretrained_model/` |

The latest step (`010000`) is also copied to the **repository root** for direct loading.

## Usage

```python
from lerobot.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("pravsels/act_block_stack")
```