sreetz-nv/so101_vials_actuators_hand_tuned

SO-101 vials-to-rack — hand-tuned actuator gains (no SysID)

This is the non-SysID arm. The asset's AnchorBench system-identification fit is overridden with the per-joint gains the workshop hand-tuned and shipped through 2026-09-04, and every joint-friction channel is pinned to zero.

Paired counterpart: sreetz-nv/so101_vials_actuators_sysid, identical in every respect except that it resolves the SysID fit from the asset. The two were collected back to back for a controlled actuator-fit A/B; use them together or the comparison is lost.

What it is

100 scripted-success episodes of the SO-101 arm picking a vial off a mat and placing it in a yellow tube rack, recorded in Isaac Lab with Newton/MJWarp physics and the OVRtx renderer. Task instruction: "Pick up the vial and place it in the rack."

episodes	100
frames	28,982
mean frames/episode	290
fps	30
robot type	so101_follower
cameras	wrist + external, 480x640 RGB
environment	Newton-So101-Auto-Vials-State-Machine-DR-v0 (domain randomized)
collection	scripted state machine; only episodes accepted by the success termination term were kept

Actuator parameters (the variable under test)

Set explicitly per joint, overriding the asset:

joint	stiffness	damping	armature	static friction	viscous friction	effort limit	velocity limit
shoulder_pan	33.0	1.6	0.02	0	0	24	4
shoulder_lift	18.0	1.8	0.02	0	0	24	4
elbow_flex	15.0	1.6	0.02	0	0	24	4
wrist_flex	15.0	1.0	0.01	0	0	18	3
wrist_roll	10.0	0.8	0.01	0	0	18	3
gripper	15.0	1.0	0.01	0	0	12	2

Friction is pinned to zero deliberately. The hand tuning predates the asset's authored newton:friction, so leaving those channels unset would resolve the SysID friction fit (0.083 to 0.411 N.m static per joint) underneath hand-tuned drive gains — neither arm of the experiment.

The grasp force servo is paired to the 15 N.m/rad gripper drive with gain 0.15 rad/(N.s) and a 0.45 rad/s jaw rate cap.

Held constant against the paired dataset

Same USD asset (so101_new_calib_SysID) with the same Robot/Sensor/Physics variants, same collision geometry, same contact material, same wrist camera and intrinsics, same solver and nconmax=512, same task, same domain randomization. Only the drive parameters differ.

Caveats

• Sim only. Nothing here speaks to real-robot transfer.
• This arm has no prior validated training result; it is a baseline constructed for this comparison.
• The paired datasets have different episode lengths (290 against 334 mean frames), so a fixed step budget gives the two arms different epoch counts.
• The arms necessarily collect with different grasp-servo constants, because their gripper drive stiffnesses differ by 4.55x. That is a difference between the arms beyond the actuator parameters themselves.