huggingface
How to evaluate policy on real robot and sim environment
I am working on evaluating a trained policy on a real robot and in a simulated environment (Isaac Gym). However, I am uncertain about the process and communication mechanisms involved.
My questions are:
- Evaluating on a real robot:
How do I retrieve real-time observations from the real robot with Lerobot?
- Evaluating in simulation (Isaac Gym):
Can I directly evaluate my trained policy in Isaac Gym?
Evaluating a policy on real robot using LeRobot package is very straight forward. The implementation is present in lerobot/scripts/control_robot.py. You can follow the instructions in the examples directory. Make sure your robot is supported by the package. Currently, LeRobot supports: SO-100, Moss, Stretch and Aloha robots.
I am not sure about evaluation in Simulation env.
Hi！Everyone！We just opensource this！ 🚀 LeIsaac — Open-source workflow combining LeRobot x IsaacSim!
Lightwheel has developed LeIsaac, a complete and reproducible workflow that demonstrates how to fine-tune GR00T N1.5 using data collected entirely from simulation — and deploy it seamlessly on real robots!
💡 What’s included? 🏠 SimReady assets — Fully prepared kitchen environment, LeRobot SO101, oranges, and more. 🎮 LeRobot teleoperation data collection in simulation — Integrated end-to-end pipeline for data collection directly in simulation. 🧠 GR00T N1.5 fine-tuning and deployment — Train GR00T N1.5 with simulation data, then deploy and validate on real hardware.
⭐ In the demo video below, the entire process is completed in just 5 simple steps.
👉 GITHUB: https://github.com/LightwheelAI/leisaac
