| --- |
| license: apache-2.0 |
| base_model: google/gemma-4-E2B-it |
| language: en |
| tags: |
| - gguf |
| - robotics |
| - tool-calling |
| - unitree-go2 |
| - ollama |
| --- |
| |
| # go2-ft-04 |
|
|
| A fine-tune of Google's Gemma 4 E2B IT that controls a **Unitree Go2** robot dog through tool calls: it |
| walks, turns, sits, does tricks, looks around and finds objects when you ask in plain English, and |
| explains what happened. Made by AUT Project 661 (LLM Agents for Robotics). |
|
|
| It's the language model of the command console in |
| [go2-mujoco-sim](https://github.com/LLM-AGENTS-FOR-ROBOTICS-AUT/go2-mujoco-sim), which downloads this |
| file on its first start, checks its SHA-256 and runs it with a private Ollama process. Every tool call |
| the model makes passes the console's safety executor (per-call limits and LiDAR clearance checks) |
| before the robot moves. |
|
|
| | | | |
| |---|---| |
| | File | `go2-ft-04.gguf`, Q4_K_M, 3,416,119,840 bytes | |
| | SHA-256 | `15a875b610d8640e4d004d71e3888ba73a7647c4eddf9b91b4df95929498293d` | |
| | Base model | google/gemma-4-E2B-it (LoRA fine-tune, merged, then quantized to 4-bit) | |
| | Prompt and tools | The compact system prompt and tool definitions in the repo (`go2/agent.py`, `go2/tools.py`) | |
| | Result | 182 of 186 (98%) on the project's tool-calling benchmark | |
|
|
| ## Intended use |
|
|
| Research and teaching with the Go2 in simulation, and supervised tests on a real Go2 EDU through the |
| console. Not approved for production, and not for driving a robot without the console's safety executor, |
| a supervisor and an emergency stop. |
|
|
| ## Licence |
|
|
| Apache 2.0, as for Gemma 4 (see Google's [model card](https://ai.google.dev/gemma/docs/core/model_card_4)). |
|
|
