Griffin Alpha-S, LIBERO fine-tune (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.
Fine-tuned from griffinlabs/Griffin-Alpha-S on all four LIBERO
suites (spatial, object, goal, 10) with the LIBERO demonstrations, 3 epochs. The other head lives on
the fast branch of this repository.
What is baked in
- Policy type
griffin_alpha; camerasobservation.images.image(agent view) andobservation.images.image2(wrist) at 256x256; 8-D state; 7-D OSC_POSE per-step deltas;n_action_steps=10(replan every 10 of the 50 predicted steps). use_relative_actions=false: LIBERO's actions are already relative commands.- Prompt header
[embodiment: LIBERO simulated Franka Emika Panda, 1 gripper; arm control mode: eef_pose];include_proprio=true,condition_on_subtask=true,num_inference_steps=1(the prompt ends withassistant\naction:at inference; see the flow-head doc). - Normalization statistics from the LIBERO training split (quantile normalization of state and actions).
LIBERO results
Closed loop, 4 suites x 10 tasks x 10 episodes = 400 episodes per configuration, single seed, 256x256
agent-view + wrist images, n_action_steps=10, native per-step OSC deltas (use_relative_actions=false).
Produced with examples/libero/run_eval.sh in the code repository.
| suite | flow head, 1 Euler step (checkpoint default) | flow head, 10 Euler steps | FAST head |
|---|---|---|---|
| libero_spatial | 98.0 | 92.0 | 96.0 |
| libero_object | 100.0 | 100.0 | 96.0 |
| libero_goal | 95.0 | 97.0 | 97.0 |
| libero_10 | 94.0 | 95.0 | 87.0 |
| four-suite | 96.8 | 96.0 | 94.0 |
Caveats: ten episodes per task is a standard error of about 1.2 points on the four-suite number; the flow and FAST checkpoints differ in more than the head (the flow one has proprio tokens in its prompt, the FAST one does not).
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-LIBERO")
preprocessor, postprocessor = make_pre_post_processors(
policy.config, pretrained_path="griffinlabs/Griffin-Alpha-S-LIBERO",
preprocessor_overrides={"device_processor": {"device": "cuda"}},
)
Fine-tune with lerobot-train --policy.path=griffinlabs/Griffin-Alpha-S-LIBERO --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.