---
language:
  - en
license: apache-2.0
base_model: google/functiongemma-270m-it
tags:
  - robotics
  - function-calling
  - gemma
  - lora
  - fine-tuned
  - edge-ai
  - jetson
pipeline_tag: text-generation
library_name: transformers
---

# FunctionGemma Robot Actions

A fine-tuned [FunctionGemma 270M](https://huggingface.co/google/functiongemma-270m-it) model that converts natural language into structured robot action and emotion function calls. Designed for real-time inference on edge devices like the NVIDIA Jetson AGX Thor.

## Overview

This model takes a user's voice or text input and outputs two function calls:

- **`robot_action`** — a physical action for the robot to perform
- **`show_emotion`** — an emotion to display on the robot's avatar screen (Rive animations)

General conversation defaults to `stand_still` with a contextually appropriate emotion.

## Example

```
Input:  "Can you shake hands with me?"
Output: robot_action(action_name="shake_hand") + show_emotion(emotion="happy")

Input:  "What is that?"
Output: robot_action(action_name="stand_still") + show_emotion(emotion="confused")

Input:  "I feel sad"
Output: robot_action(action_name="stand_still") + show_emotion(emotion="sad")
```

## Supported Actions

| Action | Description |
|--------|-------------|
| `shake_hand` | Handshake gesture |
| `face_wave` | Wave hello |
| `hands_up` | Raise both hands |
| `stand_still` | Stay idle (default for general conversation) |
| `show_hand` | Show open hand |

## Supported Emotions

| Emotion | Animation |
|---------|-----------|
| `happy` | Happy.riv |
| `sad` | Sad.riv |
| `excited` | Excited.riv |
| `confused` | Confused.riv |
| `curious` | Curious.riv |
| `think` | Think.riv |

## Performance on NVIDIA Jetson AGX Thor

Benchmarked with constrained decoding (2 forward passes instead of 33 autoregressive steps):

| Metric | Value |
|--------|-------|
| Min latency | 52 ms |
| Max latency | 72 ms |
| **Avg latency** | **59 ms** |


## Training Details

| Parameter | Value |
|-----------|-------|
| Base model | `google/functiongemma-270m-it` |
| Method | LoRA (rank 8, alpha 16) |
| Training data | 545 examples (490 train / 55 eval) |
| Epochs | 5 |
| Learning rate | 2e-4 |
| Batch size | 2 (effective 4 with gradient accumulation) |
| Max sequence length | 512 |
| Precision | bf16 |

### Quick Start

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "OpenmindAGI/functiongemma-robot-actions",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("OpenmindAGI/functiongemma-robot-actions")
model.eval()
```

## Citation

```bibtex
@misc{openmindagi-functiongemma-robot-actions,
  title={FunctionGemma Robot Actions},
  author={OpenmindAGI},
  year={2025},
  url={https://huggingface.co/OpenmindAGI/functiongemma-robot-actions}
}
```