Griffin Alpha-S base (flow-matching head)

The default head of Griffin Alpha-S: a 910M-parameter flow-matching action expert attached layer-by-layer (mixture-of-transformers) to a Qwen3-VL-4B backbone that was pre-trained on a multi-embodiment robot-data mixture. Inference integrates 10 Euler steps from noise to a 50-step action chunk.

This is the pre-trained base, meant as the starting point for fine-tuning on your own robot. For a ready-to-run policy see griffinlabs/Griffin-Alpha-S-LIBERO. The other head lives on the fast branch of this repository.

What is baked in

• Policy type griffin_alpha; 50-step action chunk; canonical 32-dimensional action vector (each embodiment's actions occupy the leading slots, the rest is zero padding; the loss masks padding per sample).
• Relative actions ON (use_relative_actions=true, relative_exclude_joints=["gripper"]): arm dimensions are predicted relative to the current observation.state, grippers absolute. Your dataset needs action feature names for this to work; see the fine-tuning guide.
• No cameras, prompt or normalization statistics baked in: image_keys is empty (all visual features of the dataset, in order, primary camera first) and the normalizer holds placeholder stats that lerobot-train / scripts/make_finetune_base.py replace from your dataset. arm_control_mode must be provided (eef_pose, joint, ...).
• Prompt: [embodiment: <text>; arm control mode: <mode>] <proprio tokens> <images> <task>\npredict subtask: ....

The expert in this base was pre-trained on a frozen backbone. The config ships freeze_backbone=false (joint training of backbone and expert) because that is the recommended fine-tuning setting; pass --policy.freeze_backbone=true to train the expert alone.

Use

Install the plugin, then any lerobot CLI understands the policy type (griffin_alpha):

pip install git+https://github.com/griffinlabs-ai/alpha-s.git

import lerobot_policy_griffin_alpha  # registers the policy types
from lerobot.policies.factory import make_pre_post_processors
from lerobot_policy_griffin_alpha import GriffinAlphaPolicy

policy = GriffinAlphaPolicy.from_pretrained("griffinlabs/Griffin-Alpha-S")
preprocessor, postprocessor = make_pre_post_processors(
    policy.config, pretrained_path="griffinlabs/Griffin-Alpha-S",
    preprocessor_overrides={"device_processor": {"device": "cuda"}},
)

Fine-tune with lerobot-train --policy.path=griffinlabs/Griffin-Alpha-S --dataset.repo_id=...; see docs/finetuning.md in the code repository, including how to rebuild the processors for a different robot or camera set.

License

Weights: CC BY-NC-SA 4.0 (see LICENSE). The plugin code is Apache-2.0. The base model, Qwen3-VL-4B-Instruct, is Apache-2.0.