Hello,
Thank you for the excellent work on MuJoCo Playground and for open-sourcing the full framework — it has been very helpful!
I am trying to reproduce the real-world cube reorientation task described in Appendix C.4 (LEAP Hand). Following the system identification and mapping procedure in C.42, and using the simulation gains kp_sim = 3.0 Nm/rad and kd_sim = 0.2 Nms/rad together with the 0.21 Nm torque limit, I derived the following DYNAMIXEL table values from the official XC330-M288-T manual:
Position P Gain (kP / address 84): 1139
Position D Gain (kD / address 80): 148
Goal Current limit (curr_lim / address 102): 406 mA
When I deploy these exact values (plus the 20 Hz policy rate), the hand behaves very differently from the videos in the paper — it is noticeably faster.
Could you please clarify a few points?
What final kP, kD, and curr_lim values (or any other control-table parameters) did you use on the real LEAP Hand hardware?
Did you set any additional limits (e.g. Velocity Limit, Profile Velocity, Drive Mode, or other parameters) that are not mentioned in the appendix?
Were there any other modifications to leaphand_node.py or the servo configuration beyond what is described in C.43?
Any insight would be greatly appreciated. I’m happy to share my exact launch file / node parameters if that helps.
Best regards,
Luis Felipe Casas
Hello,
Thank you for the excellent work on MuJoCo Playground and for open-sourcing the full framework — it has been very helpful!
I am trying to reproduce the real-world cube reorientation task described in Appendix C.4 (LEAP Hand). Following the system identification and mapping procedure in C.42, and using the simulation gains kp_sim = 3.0 Nm/rad and kd_sim = 0.2 Nms/rad together with the 0.21 Nm torque limit, I derived the following DYNAMIXEL table values from the official XC330-M288-T manual:
Position P Gain (kP / address 84): 1139
Position D Gain (kD / address 80): 148
Goal Current limit (curr_lim / address 102): 406 mA
When I deploy these exact values (plus the 20 Hz policy rate), the hand behaves very differently from the videos in the paper — it is noticeably faster.
Could you please clarify a few points?
What final kP, kD, and curr_lim values (or any other control-table parameters) did you use on the real LEAP Hand hardware?
Did you set any additional limits (e.g. Velocity Limit, Profile Velocity, Drive Mode, or other parameters) that are not mentioned in the appendix?
Were there any other modifications to leaphand_node.py or the servo configuration beyond what is described in C.43?
Any insight would be greatly appreciated. I’m happy to share my exact launch file / node parameters if that helps.
Best regards,
Luis Felipe Casas