---
license: apache-2.0
tags:
- robotics
- manipulation
- lerobot
- groot
- nvidia
- lora
- pick-and-place
base_model: nvidia/GR00T-N1.5-3B
datasets:
- so101-safe-worker
pipeline_tag: robotics
---

# GR00T Pick and Place Cube v1

A fine-tuned **NVIDIA GR00T N1.5** model for robotic pick-and-place manipulation tasks.

## Model Description

This model was fine-tuned using **LoRA** (Low-Rank Adaptation) on the SO-101 robot arm dataset for cube pick-and-place tasks.

### Training Details

| Parameter | Value |
|-----------|-------|
| Base Model | nvidia/GR00T-N1.5-3B |
| Fine-tuning Method | LoRA |
| LoRA Rank | 64 |
| LoRA Alpha | 16 |
| Training Steps | 50,000 |
| Batch Size | 8 |
| Dataset | 21,557 episodes / 1.9M frames |
| Task | Pick up cube and place in bin |
| Cameras | Front + Wrist (128x128) |
| Action Space | 4D (x, y, z, gripper) |

### Performance

- **Training Loss**: 1.17 → 0.12 (90% reduction)
- **Evaluation Success Rate**: ~60% (with proper action unnormalization)

## Usage

```python
from lerobot.policies.groot.modeling_groot import GrootPolicy

# Load the model
policy = GrootPolicy.from_pretrained("gpudad/groot-pick-place-cube-v1")
policy.to("cuda")
policy.eval()

# Use for inference
action = policy.select_action(observation_batch)
```

### Important: Action Unnormalization

The model outputs actions in normalized [-1, 1] space. For the SO-101 robot:
- XYZ: [-1, 1] 
- Gripper: needs mapping from [-1, 1] to [0, 2]

```python
# Unnormalize actions
action_min = torch.tensor([-1, -1, -1, 0])
action_max = torch.tensor([1, 1, 1, 2])
unnormalized = (action + 1) / 2 * (action_max - action_min) + action_min
```

## Framework

Trained using [LeRobot](https://github.com/huggingface/lerobot) 🤖

## License

Apache 2.0 (same as base GR00T model)