roco_smolvla

jonhpark/roco_smolvla

roco_smolvla is a machine learning model from jonhpark. Use it for the machine learning task on the model card, and read the license before you ship it in a product.

This model is a fine-tuned version of lerobot/smolvlabase on the yjsm1203/roco2 dataset for gearbox assembly task.

Downloads · 30 days

12% of all-time downloads

All-time downloads

Public

Parameters

450M

3.6 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors3.6 GB · 100%

Parameter types

How the weights are stored.

BF16447M · 99%

Model card

From the Hugging Face model README

SmolVLA Fine-tuned on roco_2 (Gearbox Assembly)

This model is a fine-tuned version of lerobot/smolvla_base on the yjsm1203/roco_2 dataset for gearbox assembly task.

Model Details

Item	Value
Base Model	`lerobot/smolvla_base`
Dataset	`yjsm1203/roco_2`
Task	Gearbox Assembly
Parameters	~450M (100M trainable)
Training Steps	20,000
Final Loss	0.006

Training Configuration

Batch Size: 64
Learning Rate: 1e-4 (with warmup and cosine decay)
GPU: RTX 5090 (32GB VRAM)
Training Time: ~6.5 hours

Training Command

lerobot-train \
  --policy.path=lerobot/smolvla_base \
  --dataset.repo_id=yjsm1203/roco_2 \
  --dataset.root=./data/roco_2 \
  --batch_size=64 \
  --steps=20000 \
  --output_dir=outputs/train/roco_smolvla \
  --job_name=roco_smolvla_gearbox \
  --policy.device=cuda \
  --save_freq=5000 \
  --log_freq=100 \
  --wandb.enable=false \
  --policy.repo_id=jonhpark/roco_smolvla \
  --rename_map='{"observation.images.head": "observation.images.camera1", "observation.images.left_hand": "observation.images.camera2", "observation.images.right_hand": "observation.images.camera3"}'

Task Description

Assemble the gearbox by placing each small gear onto the tree pins one at a time.
Install the last small gear in the center, then put the cover on.

Dataset Features

observation.images.head: (3, 240, 320) - Head camera
observation.images.left_hand: (3, 240, 320) - Left hand camera
observation.images.right_hand: (3, 240, 320) - Right hand camera
observation.state: (28,) - Robot state
action: (14,) - Robot action

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("jonhpark/roco_smolvla")
policy.to("cuda")

# Use with your robot or evaluation environment

Training Metrics

Step	Loss	Gradient Norm
100	1.149	4.229
1,000	0.067	0.565
5,000	0.019	0.334
10,000	0.010	0.279
15,000	0.006	0.231
20,000	0.006	0.185

Checkpoints

Available checkpoints in this repository:

checkpoints/005000/ - Step 5,000
checkpoints/010000/ - Step 10,000
checkpoints/015000/ - Step 15,000
checkpoints/020000/ - Step 20,000 (Final)

License

Apache 2.0

Created: Jan 5, 2026
Updated: Jan 7, 2026