SmolVLA fine tuned on OpenArm Lego pick and place (n50)

This is a SmolVLA policy fine tuned on 50 teleoperated demonstrations of a single arm pick and place task on the OpenArm. It was trained by the OPENROB project at FURG as the 50 episode condition of a data scale study.

Task

The right arm picks a single white Lego brick from the table and places it inside a container at a fixed position. The brick position and orientation vary between episodes.

Instruction: put the lego brick in the box

Training data

Dataset: openrob-furg/openarm-single-pick-place-lego

This model was trained on episodes 0 to 49 of that dataset (13,104 frames at 30 fps).

Inputs and outputs

	Key	Description
Input	`observation.images.cam_chest`	RGB, Intel RealSense D435i on the chest
Input	`observation.images.right_cam_wrist`	RGB, Intel RealSense D405 on the right wrist
Input	`observation.state`	8 values: right arm joints 1 to 7 and the gripper
Output	`action`	Chunks of 50 absolute joint position targets (8 values each)

The camera names are mapped to the SmolVLA image slots by the saved preprocessor, so the policy expects the keys above.

Training setup

Setting	Value
Base model	`lerobot/smolvla_base` (0.45B parameters)
Trained parameters	Action expert and state projection; vision encoder and language model frozen
Steps / batch size	20,000 / 32
Optimizer	AdamW, peak learning rate 1e-4, cosine decay to 2.5e-6, 666 warmup steps
Seed	1000
Hardware	1x NVIDIA RTX 4070 Ti (12 GB), about 3 hours
Software	LeRobot 0.6.1

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("openrob-furg/smolvla-openarm-single-pick-place-lego-n50")

Use the same LeRobot version (0.6.1) for inference to avoid configuration mismatches.

Related models

Model	Training episodes
openrob-furg/smolvla-openarm-single-pick-place-lego-n50	50
openrob-furg/smolvla-openarm-single-pick-place-lego-n100	100
openrob-furg/smolvla-openarm-single-pick-place-lego-n200	200