Policy Inference

---

1. Policy Inference

Run inference with a trained policy model and automatically record the results:

/isaac-sim/python.sh src/lerobot/scripts/lerobot_record.py \
    --robot.type=walker_s2_sim \
    --robot.headless=false \
    --task=Part_Sorting \
    --policy.path=your_checkpoint_dir \
    --dataset.repo_id=your_org/eval_Part_Sorting \
    --dataset.single_task="Part Sorting" \
    --dataset.num_episodes=2 \
    --dataset.push_to_hub=false \
    --dataset.episode_time_s=100000000 \
    --dataset.num_image_writer_processes=4 \
    --dataset.root=${workspaceFolder}/datasets/Part_Sorting/your_policy_type \
    --dataset.video=true \
    --play_sounds=false

Argument	Description	Default / Notes
--robot.type	Robot type	walker_s2_sim
--robot.headless	Whether to run in headless mode	false
--task	Task name (Part_Sorting, Conveyor_Sorting, Foam_Inlaying, Packing_Box)	Foam_Inlaying
--policy.path	Policy checkpoint path (local or Hugging Face)	Required
--dataset.repo_id	Dataset identifier; in inference it should follow xx/eval_xx	Required
--dataset.single_task	Task description	Required (or auto-set by task)
--dataset.num_episodes	Number of inference episodes	50
--dataset.root	Output path for inference results	Required
--dataset.video	Whether to record video	true
--dataset.fps	FPS	30
--dataset.episode_time_s	Episode duration (seconds)	60
--dataset.push_to_hub	Upload to Hugging Face	false
--play_sounds	Play audio prompts	true

Notes:

1. Policy inference uses the lerobot_record.py script. Specify the policy model path via --policy.path. The script will load the model, run inference in the simulation environment, and record the resulting data automatically.
2. In inference mode, --dataset.repo_id should use the xx/eval_xx format (e.g., your_org/eval_Part_Sorting) to distinguish it from dataset naming in data-collection mode.
3. Replace ${workspaceFolder} with your actual working directory path.

---