| --- |
| license: cc-by-nc-4.0 |
| library_name: lerobot |
| pipeline_tag: robotics |
| base_model: Adaptive-Machine-Patterns/amp-tadpole |
| tags: |
| - robotics |
| - lerobot |
| - vision-language-action |
| - vla |
| - amp-toy-robotics |
| --- |
| |
| # AMP Tadpole for LeRobot |
|
|
| The weights of [AMP Tadpole](https://huggingface.co/Adaptive-Machine-Patterns/amp-tadpole) |
| v0.2.0 in LeRobot's format, for the `tadpole` policy type from |
| [lerobot-policy-tadpole](https://github.com/Adaptive-Machine-Patterns/lerobot-policy-tadpole). |
|
|
| AMP Tadpole is a small vision-language-action model (6.5M parameters) that drives a |
| six-joint arm from three fixed camera images, the arm's state and a plain-English |
| instruction, trained from scratch on the |
| [AMP Toy Robotics Benchmark](https://github.com/Adaptive-Machine-Patterns/amp-toy-robotics) |
| (93.7% of instructions on its 300 held-out scenes). Its results, training, limitations |
| and credits are on [the main model card](https://huggingface.co/Adaptive-Machine-Patterns/amp-tadpole). |
|
|
| **Try it in your browser:** the |
| [AMP Embodied AI Workbench Demo](https://huggingface.co/spaces/Adaptive-Machine-Patterns/amp-embodied-ai-workbench-demo). |
|
|
| ## Use |
|
|
| ```bash |
| pip install git+https://github.com/Adaptive-Machine-Patterns/lerobot-policy-tadpole |
| ``` |
|
|
| ```python |
| from lerobot_policy_tadpole import TadpolePolicy, make_tadpole_pre_post_processors |
| |
| policy = TadpolePolicy.from_pretrained("Adaptive-Machine-Patterns/amp-tadpole-lerobot") |
| pre, post = make_tadpole_pre_post_processors(policy.config) |
| ``` |
|
|
| or `--policy.path=Adaptive-Machine-Patterns/amp-tadpole-lerobot` with LeRobot's tools. |
|
|
| ## What is in it |
|
|
| The same network and weights as the main repository's v0.2.0 (step 330,000), and it gives |
| the same answers: checked on 36 held-out observations through LeRobot's own processors, |
| with no difference. Two parts used only in training are kept, so it can be fine-tuned: |
| the head with which each image token redraws its patch, and the "no state" token. That |
| makes 6,557,191 parameters. The word table holds the 32 tokens of the benchmark's |
| instructions, plus one shared row for any other word. |
|
|
| ## Licence |
|
|
| **Free for non-commercial use** under |
| [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) (see LICENSE and NOTICE). |
| **For commercial use**, licences are available from Adaptive Machine Patterns Ltd: |
| [adaptive-machine-patterns.com/embodied-ai-workbench](https://adaptive-machine-patterns.com/embodied-ai-workbench/) |
| or amp@adaptive-machine-patterns.com. The plugin's code is Apache-2.0. |
|
|
| **No warranty.** Provided "as is", without warranty of any kind (sections 5 and 6 of the |
| licence). Not suitable for controlling real machinery. |
|
|
| LeRobot is a Hugging Face project; its mention implies no endorsement. |
|
|
| © 2026 Adaptive Machine Patterns Ltd, registered in England & Wales, no. 15618869. |
|
|
