CoRL2026-CSI/SmolVLA-UR7e-ArrangeBlock_50epoch

This is a LeRobot SmolVLA policy fine-tuned from lerobot/smolvla_base on CoRL2026-CSI/UR7e-CaP_arrange_block_100epi_10fps.

The model is intended for UR7e ArrangeBlock manipulation experiments using RGB observations, robot proprioception, language instructions, and continuous action chunks. It is uploaded as a LeRobot policy checkpoint and should be loaded through the matching LeRobot SmolVLA implementation used for training.

Model Details

• Policy type: SmolVLA
• Base policy: lerobot/smolvla_base
• Vision-language model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
• Action chunk size: 50
• Action steps: 50
• Max state/action dims: 32 / 32
• Vision encoder frozen: true
• Train expert only: true
• Train state projection: true

Fine-Tuning Setup

• Dataset: CoRL2026-CSI/UR7e-CaP_arrange_block_100epi_10fps
• Training steps: 9250
• Approx. epochs: 50.26
• Final training samples: 2368000
• Final training loss: 0.010289
• Runtime: 6.12 hours
• Per-GPU batch size: 128
• Gradient accumulation steps: 1
• Number of GPUs: 2
• Effective batch size: 256
• Optimizer lr: 0.0001
• Optimizer betas: [0.9, 0.95]
• Weight decay: 1e-10
• Scheduler warmup/decay: 1000 / 30000
• Final decay lr: 2.5e-06
• DataLoader workers: 8
• DataLoader prefetch factor: 1

Camera Mapping

• observation.images.realsense_topview -> observation.images.camera2
• observation.images.realsense_wrist -> observation.images.camera1

Image Augmentation

• affine: RandomAffine {'degrees': [-5.0, 5.0], 'translate': [0.05, 0.05]}
• brightness: ColorJitter {'brightness': [0.8, 1.2]}
• contrast: ColorJitter {'contrast': [0.8, 1.2]}
• hue: ColorJitter {'hue': [-0.05, 0.05]}
• saturation: ColorJitter {'saturation': [0.5, 1.5]}
• sharpness: SharpnessJitter {'sharpness': [0.5, 1.5]}

Inputs

• observation.images.camera1: VISUAL, shape [3, 256, 256]
• observation.images.camera2: VISUAL, shape [3, 256, 256]
• observation.images.camera3: VISUAL, shape [3, 256, 256]
• observation.state: STATE, shape [6]

Outputs

• action: ACTION, shape [7]

Usage

Install and use the same LeRobot checkout/environment that contains the SmolVLA policy implementation, then point policy.path to this Hub repo.

lerobot-record \
  --robot.type=<your_robot> \
  --dataset.repo_id=<your_eval_dataset_repo> \
  --policy.path=CoRL2026-CSI/SmolVLA-UR7e-ArrangeBlock_50epoch \
  --episodes=10

For local Python usage, load the policy with LeRobot's policy factory from the training checkout.

Evaluation

This upload records the offline training run metrics only. No rollout success rate is claimed here unless a separate real/sim evaluation is added later.

Final logged training metrics:

• loss: 0.010289
• grad norm: 0.101402
• learning rate: 2.5000801319183248e-06
• update time: 1.1543 s/step
• dataloading time: 1.0103 s/step

Limitations and Safety

This model is a robot control policy and can produce unsafe actions if deployed on hardware without appropriate validation, workspace limits, emergency stop handling, and task-specific safety checks. Test in simulation or a constrained setup before any physical deployment.

The model is specialized to the training dataset, camera mapping, calibration, action space, and embodiment configuration. It may not transfer reliably to different robots, camera placements, object layouts, or tasks without further validation or fine-tuning.

License and Terms

The training dataset is marked apache-2.0, and the SmolVLM2 component is marked apache-2.0. The lerobot/smolvla_base model card does not currently declare a license field, so this fine-tuned model is conservatively marked as other. Users are responsible for checking the applicable base model, dataset, and deployment terms before use.

Files

• model.safetensors: fine-tuned policy weights
• config.json: LeRobot SmolVLA policy config
• train_config.json: training configuration
• policy_preprocessor.json and policy_postprocessor.json: LeRobot processor pipelines
• policy_*_processor.safetensors: normalization/statistics state used by processors