LingBot-VLA: Pick and Place Toy to Bucket

This checkpoint is a full-parameter fine-tune of LingBot-VLA-4B on the AlvinAi/pick_and_place_toy_to_bucket LeRobot dataset.

Task instruction

pick and place toy to bucket

The training pipeline formats it as <bos>pick and place toy to bucket\n before tokenization.

Dataset and robot layout

• 82 episodes, 27,962 frames, 30 FPS
• Cameras: observation.images.front and observation.images.side
• State/action dimensions:
  1. shoulder_pan.pos
  2. shoulder_lift.pos
  3. elbow_flex.pos
  4. wrist_flex.pos
  5. wrist_roll.pos
  6. gripper.pos
• Arm joints are trained as delta actions.
• The gripper is trained as an absolute action.
• The action horizon is 50 steps, with episode-tail padding excluded from both normalization statistics and the training loss.

The dataset's gripper values are approximately in the range 0–34. Hardware inference must use the same calibration and units; do not assume a 0–100 gripper range.

Fine-tuning setup

• Full-parameter fine-tuning of the VLM, vision encoder, and action expert
• 4 epochs
• Global batch size 256 on 4 GPUs
• Learning rate 2e-5, constant schedule
• FP32/TF32 FSDP2 training
• Mean/std normalization from the accompanying norm_stats.json

Included files

• Hugging Face model weights and tokenizer/processor configuration
• lingbotvla_cli.yaml: resolved training configuration
• norm_stats.json: task-specific normalization statistics
• robot_config.yaml: feature mapping used during training and inference

For inference, use norm_stats.json and robot_config.yaml from this model repository so that input padding, normalization, action restoration, and the absolute gripper output remain aligned with training.