Beverage SmolVLA (40,000 steps)

SmolVLA fine-tuned from lerobot/smolvla_base for bimanual beverage preparation.

Training run: 20261103_024412_232249 (Asia/Seoul).
Checkpoint step: 40,000.
Selected task IDs: 0, 1, 2, 4 (stir, pour water, pour green tea latte powder, pour coffee powder).
Robot state and action: 12 dimensions, left arm followed by right arm.
Cameras: top, left wrist, right wrist; mapped to camera1, camera2, camera3.
Action chunk size: 50.

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
from lerobot.policies import make_pre_post_processors

repo_id = "2chan/beverage-smolvla-40k"
policy = SmolVLAPolicy.from_pretrained(repo_id)
preprocessor, postprocessor = make_pre_post_processors(
    policy.config, pretrained_path=repo_id
)

Pass camera observations, robot state, and a task string through the preprocessor before action selection. Apply the postprocessor to recover actions in the original scale.

This repository contains the pretrained model, configuration, and normalization processors. Optimizer/scheduler/RNG state is not included. This upload does not establish physical robot performance.