THU98/smolvla_franka_5task_80k_gpu2

SmolVLA Franka 5-Task 80K

This is a task-specific SmolVLA policy fine-tuned from lerobot/smolvla_base for five language-conditioned bimanual Franka manipulation tasks.

Model details

• Architecture: SmolVLA (approximately 450M parameters)
• Fine-tuned components: action expert and state projection (approximately 100M trainable parameters)
• Observation: three RGB views, a language instruction, and a 16-dimensional robot state
• Output: a 50-step action chunk with 16 action dimensions per step
• Base vision-language model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
• Framework: LeRobot

The three camera streams used during training were mapped as follows:

Source field	Policy field
observation.images.cam_high	observation.images.camera1
observation.images.cam_left_wrist	observation.images.camera2
observation.images.cam_right_wrist	observation.images.camera3

Tasks

The training set contains 50 demonstrations for each task (250 episodes in total):

1. Place the green bowl on top of the cabinet with the left arm.
2. Place the green bowl on the blue plate with the left arm.
3. Place the gray bowl on the green plate with the left arm.
4. Place the cup on the tray with the left arm.
5. Place the tape in the green basket with the left arm.

Training data

The policy was trained on a local LeRobot v3 dataset identified as thu98/smolvla_franka_5task_train.

• Episodes: 250
• Frames: 58,142
• Frequency: 20 FPS
• Independent validation split: none; all 250 episodes were used for training

Training configuration

Setting	Value
Training steps	80,000
Batch size	16
Precision	bfloat16 training autocast
Optimizer	AdamW
Peak learning rate	1e-4
Warmup	1,000 steps
Decay	cosine, 30,000 steps to 2.5e-6
Weight decay	1e-10
Gradient clipping	10
Image augmentation	disabled
Checkpoint interval	1,000 steps
Training GPU	one NVIDIA GPU

The final logged training loss was approximately 0.031. This is an optimization metric and should not be interpreted as task success rate.

Usage

Install a compatible version of lerobot, then load the policy with:

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained(
    "THU98/smolvla_franka_5task_80k_gpu2"
)

At inference time, observations must follow the same camera mapping, robot-state ordering, action ordering, and language-instruction convention used for training. The 16 state/action dimensions are ordered as seven left-arm joints, left gripper closedness, seven right-arm joints, and right gripper closedness.

Evaluation and limitations

This checkpoint has not been evaluated on a held-out validation set or reported with real-robot task success rates. It is specialized for the training robot, camera placement, scene layout, objects, and instruction phrasing. Performance under distribution shift is unknown. Validate safety constraints and task success in a controlled environment before real-robot deployment.

Citation

This checkpoint is derived from SmolVLA. Please refer to the base model card and the LeRobot project for the appropriate citation and attribution.