RLinf-Pi05-RealWorld-DualYAM-Clean_Desk

Pi0.5 checkpoint used for the dual-arm YAM tabletop cleanup demo in RPent.

• Inputs: top, left-wrist and right-wrist RGB images, a 14-dimensional robot state, and a task instruction.
• Actions: 30-step chunks with 14 dimensions: six joint angles and one gripper value per arm. Joint actions use deltas; gripper values are absolute, with 0 closed and 1 open.
• Model: PaliGemma 2B and Gemma 300M action expert, with a 32-dimensional internal action representation.
• Weights: FP32 model.safetensors, exported from the deployment checkpoint without changing tensor values. norm_stats.json contains its matching normalization statistics.

Use an RLinf/RPent environment configured for pi05_yam_joint; see the YAM setup guide.

hf download RLinf/RLinf-Pi05-RealWorld-DualYAM-Clean_Desk \
  --local-dir ./yam-clean-desk

python -m rpent.robots.components.pi05_vla_server \
  --embodiment yam \
  --model-path ./yam-clean-desk \
  --norm-stats-path ./yam-clean-desk/norm_stats.json \
  --transport socket --host 127.0.0.1 --port 8220

The checkpoint is specific to the demonstration setup. Camera placement, robot calibration and task data must match the deployment setup.