LeRobot Dataset Replay in Simulation

🎯 Objective

Re-simulate episodes from the LeRobot dataset where the robot picks up a golf ball and places it in a cup.

🧭 Workflow

1. 3D Scene Generation

• Use Hunyan3D-2.0 with a single video frame to generate a 3D .glb model of the scene.
• Process the GLB:
  • Decompose into components.
  • Filter meshes by object type.
  • Scale each mesh using bounding box to match real-world dimensions.
  • Rotate for correct object orientation.

2. Key Frame Identification

• Identify two key frames:
  • Pick: Gripper holds the ball.
  • Place: Ball is dropped into the cup.
• Use motor signals:
  • Difference between the mobile jaw (Follower) and teleop handle (Leader).
  • When gripping
    • Follower: mobile jaw remains fixed around the ball
    • Leader: holding loop controlling the gripper is completely closed

3. Pose Estimation

• Use Forward Kinematics to compute gripper's Cartesian pose at pick/place key frames.
• Apply a fixed translation vector from the link center to the tip of the gripper's fixed jaw.
• This defines the initial ball and cup positions in sim.

4. Scene Setup & Replay

• Add the ball and cup at their computed locations.
• Replay the episode in simulation.
• Evaluate success:
  • Check if the golf ball ends up inside the cup.
  • Use XY distance between ball and cup at final frame.

5. Evaluation

• Replayed all episodes in the dataset.
• Achieved 42% success rate.

⚠️ Issues & Observations

• Discrepancy observed when gripper picks the ball horizontally.
  • Possibly due to torque effects when the arm is stretched, shifting the center of mass.
• Dataset likely logs goal position, but simulator should use the present position of the Feetech motor.
• Applying a scaling factor of ~0.9 to gripper position:
  • Helps with pick alignment.
  • Causes errors during lifting in low-torque configurations.
• Also needed to scale gripper motor signals to squeeze the ball

Future Work

1. improve the success rate of the episode replay script to ~ 100%. Adding the actual motor position to the dataset samples is required. The current assumption is that it contains the control position, which cause the simulator to diverge when the arm is stretched out.
2. generate a pure sim dataset, using the sim camera recording but the robot state from the dataset
3. train any policy on this sim dataset
4. evaluate the policy in sim
5. transfer sim policy to real camera

Caveat

It is not clear how to transfer SIM2REAL since the policy is fined-tuned on the sim camera, not the actual feed.