We share the process of optimizing acceleration feedforward control for a 6-axis robot arm. By using planner acceleration + LPF instead of numerical differentiation, we achieved 21.2% improvement in position RMSE and torque stability simultaneously.
A robot arm that is fed position targets moves exactly as well as the host keeps time. We put our hard real-time layer on a Jetson AGX Orin in front of an unmodified UFACTORY xArm 7 used for humanoid teleoperation. Send jitter stayed within 5 µs, the box's 6.5 ms lag was cancelled, and the arm went from 1.57 s behind the plan to on time. Tool-tip error at the same instant fell from 165 mm to 0.3 mm, path error from 1.11 mm RMS to 0.008 mm, with zero box errors across 30 runs and a 1.57 kg load swinging from the gripper. This post is the whole stack, kernel to motor, with the measurements behind every layer.
We validate the effects of friction compensation and acceleration feedforward on a 6-axis collaborative robot through 160 runs of 2x2 experiments. Friction coefficients were tuned using an AI agent (ralph-loop), and validation experiments were conducted with automated scripts. Friction compensation alone achieved 36% improvement (Cohen's d=1.49), accounting for 93% of total improvement.