Inference Interfaces#

This page focuses on deployment and inference-related interfaces, covering Runner, Operator, and inference data processing.

Purpose#

Load trained models in real-robot or simulation environments
Read sensor inputs and perform action prediction
Send predicted actions to the robot control interface

Core Parameters#

Inference is typically controlled through the inference (real-robot) or eval (simulation evaluation) configuration.

Common `inference` Fields#

Field	Description
`type`	Inference Runner type (e.g., `AlohaInferenceRunner`)
`task_descriptions`	Task description dictionary
`dataset`	Inference input data processing
`denormalize_action`	Action denormalization configuration
`operator`	Robot communication interface and ROS topic configuration

Common `eval` Fields#

Field	Description
`type`	Evaluation Runner type (e.g., `LiberoEvalRunner`)
`task_suite_name`	Task suite
`dataset`	Evaluation data processing pipeline
`denormalize_action`	Action denormalization configuration

Minimal Example#

WANDB_MODE=disabled HF_ENDPOINT=https://hf-mirror.com \
python scripts/inference_real_robot.py \
  --config configs/gr00t/gr00t_eagle_3b_aloha_fold_towel_3cam_4090_full_train.py \
  --ckpt-path ./work_dirs/.../checkpoints/step-xxxx.pt

Platform Entry Points#

Inference deployment overview: ../tutorials/inference/index.rst
Aloha: ../tutorials/inference/Aloha.md
Tron2: ../tutorials/inference/Tron2.md
UR3: ../tutorials/inference/UR3.md