I need your help! I can't training ACT policy as good as you...

[mamimaka]

Hi, I've recently been using your ACT strategy on my platform for imitation learning. Here's the setup:

• Environment: CoppeliaSim 4.5.2 (simulation environment)
• Robotic arm: single UR3 with gripper(7 DOF)
• Task: grabbing and placing wooden blocks at a destination (pick and place)
  During training, I've found that more than just using 50 or 100 demonstrations is needed to get the robot to complete the task, even though my task is simpler compared to what's in the paper. Specifically, after training with 50 demonstrations, the ACT algorithm couldn't get my robotic arm to reach around the wooden block, and with 100 demonstrations, it couldn't successfully grab the block. Here are my hyperparameters:
• policy_class: ACT
• kl_weight: 10
• chunk_size: 100
• hidden_dim: 512
• batch_size: 2
• dim_feedforward: 3200
• num_epochs: 2000
  I want your help on this.