Skip to main content
Put one checkpoint at the Hugging Face repository root:
  • JAX/Orbax: contents of one OpenPI training step: params/ and exactly one assets/<id>/norm_stats.json. For non-default inference settings, add openpi_config.json at the root.
  • OpenPI PyTorch: contents of the conversion output: config.json, model.safetensors, and assets/<id>/norm_stats.json. Import with --openpi-config-name NAME using a config registered by Servo.
  • LeRobot: contents of one checkpoint’s pretrained_model/ directory, following LeRobot’s checkpoint layout: config.json, model.safetensors, policy_preprocessor.json, policy_preprocessor_step_3_normalizer_processor.safetensors, policy_postprocessor.json, and policy_postprocessor_step_0_unnormalizer_processor.safetensors.