This repository facilitates imitation learning on HEBI arm platforms using the LeRobot framework.

• HEBI Organization: huggingface.co/hebirobotics
• Models & Datasets: Available via Hariharan’s profile
• Framework Documentation: LeRobot Docs

You can visualize recorded datasets directly through the Hugging Face GUI.

• Tool: LeRobot Dataset Visualizer
• Example: HEBI Arm Pick and Drop - Episode 0

Core scripts for recording, training, and evaluation are located in src/lerobot/scripts.

The hardware integration logic is located in src/lerobot/robots/hebi_arm. A key feature of the HebiConfig is the send_action boolean:

• send_action: False (Default): Recommended for data collection. The system records feedback from the arm without sending commands. This allows you to control the arm using external scripts (e.g., the AR Kit Python demo) while LeRobot captures the state.
• send_action: True: Required for evaluation. This allows the trained policy to send commands directly to the arm to execute learned behaviors.

To ensure the LeRobot framework recognizes the HEBI platform, you must manually register the module in the robot factory scripts.

1. Update Imports: Locate the file where robots are initialized (For example, line 92 in src/lerobot/scripts/lerobot_record.py). Find the import block and add hebi_arm to the end:

   from lerobot.robots import (
       ...,
       hebi_arm,
   )

2. Script Modification: After updating the import, ensure your script references the hebi_arm configuration. This allows the make_robot() function to correctly map your YAML config to the HEBI hardware classes.

• Tutorial: Imitation Learning on Robots
• Data Collection Tips: What makes a good dataset?