cycle=003 sample=00 | first_delta=[+0.060, -0.254, -0.059, -0.382, -0.251, +0.166, +1.876] | last_delta=[+0.205, -0.144, -0.081, -0.233, -0.211, +0.270, +1.754] | max_abs_delta=[+0.205, +0.554, +0.179, +0.382, +0.416, +0.374, +2.119] | span=[+0.160, +0.411, +0.312, +0.273, +0.270, +0.281, +0.852] | mean_step=[+0.043, +0.084, +0.074, +0.046, +0.058, +0.078, +0.289]
cycle=003 sample=01 | first_delta=[+0.022, -0.180, +0.049, -0.557, -0.352, +0.022, +1.876] | last_delta=[+0.085, -0.261, +0.101, -0.150, -0.262, +0.014, +2.119] | max_abs_delta=[+0.129, +0.326, +0.189, +0.563, +0.393, +0.212, +2.606] | span=[+0.190, +0.529, +0.298, +0.413, +0.265, +0.229, +1.095] | mean_step=[+0.039, +0.122, +0.088, +0.079, +0.059, +0.047, +0.284]
cycle=003 sample=02 | first_delta=[+0.129, -0.072, +0.006, -0.341, -0.247, +0.044, +1.511] | last_delta=[+0.115, -0.059, +0.178, -0.060, -0.544, -0.058, +1.632] | max_abs_delta=[+0.326, +0.235, +0.250, +0.341, +0.550, +0.221, +1.998] | span=[+0.222, +0.361, +0.380, +0.281, +0.304, +0.279, +0.974] | mean_step=[+0.054, +0.099, +0.074, +0.055, +0.049, +0.055, +0.289]
cycle=003 sample=03 | first_delta=[+0.005, -0.094, +0.098, -0.541, -0.377, -0.006, +2.363] | last_delta=[+0.161, +2.916, -0.854, +0.196, -0.217, -0.145, +2.241] | max_abs_delta=[+0.161, +2.916, +0.854, +0.541, +0.460, +0.214, +2.485] | span=[+0.224, +3.418, +1.172, +0.831, +0.271, +0.345, +0.609] | mean_step=[+0.035, +0.261, +0.115, +0.066, +0.057, +0.056, +0.178]
Under the same inference and training scenarios, why is the action delta value output based on observations so small when performing inference on the actual robot, causing the robotic arm to vibrate in place during real-world operation? Additionally, the dataset has been normalized, and the training loss is on the order of 0.01. What could be the reason for the extremely small output actions on the real robot, preventing the robotic arm from effectively completing the task?
cycle=003 sample=00 | first_delta=[+0.060, -0.254, -0.059, -0.382, -0.251, +0.166, +1.876] | last_delta=[+0.205, -0.144, -0.081, -0.233, -0.211, +0.270, +1.754] | max_abs_delta=[+0.205, +0.554, +0.179, +0.382, +0.416, +0.374, +2.119] | span=[+0.160, +0.411, +0.312, +0.273, +0.270, +0.281, +0.852] | mean_step=[+0.043, +0.084, +0.074, +0.046, +0.058, +0.078, +0.289]
cycle=003 sample=01 | first_delta=[+0.022, -0.180, +0.049, -0.557, -0.352, +0.022, +1.876] | last_delta=[+0.085, -0.261, +0.101, -0.150, -0.262, +0.014, +2.119] | max_abs_delta=[+0.129, +0.326, +0.189, +0.563, +0.393, +0.212, +2.606] | span=[+0.190, +0.529, +0.298, +0.413, +0.265, +0.229, +1.095] | mean_step=[+0.039, +0.122, +0.088, +0.079, +0.059, +0.047, +0.284]
cycle=003 sample=02 | first_delta=[+0.129, -0.072, +0.006, -0.341, -0.247, +0.044, +1.511] | last_delta=[+0.115, -0.059, +0.178, -0.060, -0.544, -0.058, +1.632] | max_abs_delta=[+0.326, +0.235, +0.250, +0.341, +0.550, +0.221, +1.998] | span=[+0.222, +0.361, +0.380, +0.281, +0.304, +0.279, +0.974] | mean_step=[+0.054, +0.099, +0.074, +0.055, +0.049, +0.055, +0.289]
cycle=003 sample=03 | first_delta=[+0.005, -0.094, +0.098, -0.541, -0.377, -0.006, +2.363] | last_delta=[+0.161, +2.916, -0.854, +0.196, -0.217, -0.145, +2.241] | max_abs_delta=[+0.161, +2.916, +0.854, +0.541, +0.460, +0.214, +2.485] | span=[+0.224, +3.418, +1.172, +0.831, +0.271, +0.345, +0.609] | mean_step=[+0.035, +0.261, +0.115, +0.066, +0.057, +0.056, +0.178]
Under the same inference and training scenarios, why is the action delta value output based on observations so small when performing inference on the actual robot, causing the robotic arm to vibrate in place during real-world operation? Additionally, the dataset has been normalized, and the training loss is on the order of 0.01. What could be the reason for the extremely small output actions on the real robot, preventing the robotic arm from effectively completing the task?