3 ms·
One would expect it to at least have calibration and inverse kinematics solved. That's still a massive pain point for the SO-ARM101 despite it being so common n
by moffkalast 18d ago
One would expect it to at least have calibration and inverse kinematics solved. That's still a massive pain point for the SO-ARM101 despite it being so common now, since there's no base_link to gripper calibration nor sag compensation and every center offset stacks.
Lerobot banked on having full DL control of each joint so they wouldn't have to do that properly, but now VLAs just use inverse kinematics as a crutch and predict gripper offsets, so we're back to square one.
- drapado 18d agoHi! do you have any references for VLAs using inverse kinematics? I'm very curious and can't find much
- moffkalast 18d agoI'm not super up to date on new releases, the one I'm most familiar with is X-VLA which predicts gripper poses and relative movement deltas for kinematics to move to instead of joints, and trains on forward kinematic gripper poses. I think Intern and Octo have the same approach.