HardenZhao/pi05-stack-cubes-umi-step20000

Pi0.5 UMI Single-Franka — SFT Step 20000

This is an RLinf deployment export of the global_step_20000 SFT checkpoint for the UMI single-Franka stack_cubes_umi_1_franka task.

The checkpoint was converted from the RLinf FSDP training format to the bare OpenPI PyTorch format expected by the RLinf real-world evaluator. All floating point tensors are stored as BF16.

Files

• model.safetensors: Pi0.5 model weights.
• config.json: Pi0.5 model architecture configuration.
• norm_stats/norm_stats.json: task normalization statistics.

Download

hf download <OWNER>/<REPOSITORY> --local-dir /path/to/pi05_step_20000

For a private repository, authenticate first with hf auth login using an account that has read access.

RLinf deployment

From the RLinf repository root:

bash evaluations/run_umi_left_franka_eval.sh \
  --model_path /path/to/pi05_step_20000 \
  --norm_stats_path /path/to/pi05_step_20000/norm_stats \
  --wrist_camera <WRIST_CAMERA_SERIAL> \
  --front_camera <FRONT_CAMERA_SERIAL>

The evaluator also requires the Franka and OpenPI runtime environments described in the RLinf real-world evaluation documentation.

Checkpoint metadata

• Training step: 20000
• Policy: Pi0.5, single Franka, TCP rotation-6D action space
• Action dimension: 10
• Action chunk used by RLinf: 20
• Export dtype: bfloat16
• Number of tensors: 667

This export is intended for deployment and evaluation. It does not contain the optimizer or dataloader state required to resume FSDP training.