simtos_3031_40k_notorso_

Pi0.5 model fine-tuned from lerobot/pi05_base on rainbowrobotics/simtos_sort_30_31_merged_notorso_qfixed.

Model Details

• Base model: lerobot/pi05_base (PaliGemma 2B + Gemma 300M action expert)
• Robot: RBY1 (dual arm)
• Task: Recyclable sorting (cans, PET bottles, paper)
• Training steps: 40,000
• Batch size: 20
• dtype: bfloat16

Input / Output

Feature	Shape	Type
observation.state	(16,)	right_arm (7) + left_arm (7) + grippers (2)
observation.images.front	(480, 640, 3)	Front camera
observation.images.right	(640, 480, 3)	Right camera
observation.images.left	(640, 480, 3)	Left camera
action	(16,)	right_arm (7) + left_arm (7) + grippers (2)

Training Config

• Learning rate: 2.5e-5 (cosine decay to 2.5e-6)
• Warmup steps: 1,000
• Optimizer: AdamW (weight_decay=0.01)
• Normalization: QUANTILES (state & action)
• freeze_vision_encoder: false
• train_expert_only: false