PID control works if plant dynamics remain stable and known. It breaks when contact changes the plant abruptly. PID controllers are simple to understand and tune and cheap enough for a microcontroller. They still suit many tasks. At Neuralzome we use them where dynamics are well characterized and the operating point is stable.
A specific class of manipulation task creates a structural limit for PID control. Understanding that limit helps determine whether a learned policy is justified or whether tuning is enough.
What PID Actually Does
A PID controller computes output from three terms: proportional response to current error, integral accumulation of past error, and derivative response to error change. Gains Kp Ki and Kd are usually tuned empirically for a particular plant and operating condition.
PID works when plant dynamics are roughly linear and stable across the operating range. Gain tuning implicitly models the plant. Kp assumes a particular link between control effort and output response. When that link changes controller behavior degrades.
A conveyor motor holding target speed has stable dynamics. A PID tuned once can run reliably for years. Load changes usually remain within the range that keeps its gains suitable.
When Dynamics Change
Contact uncertainty breaks PID assumptions. In peg in hole insertion the arm moves freely before contact with low and known inertia. Contact abruptly adds end effector compliance fixture stiffness and friction at the peg hole interface. Those parameters are absent from the free space model.
PID can be tuned for contact or free space but not both at once. Aggressive free space gains cause contact overshoot which may bend the peg or trigger a safety fault. Conservative contact gains make approach sluggish and hurt cycle time.
Control engineers often use gain scheduling with separate controllers for each phase plus transition detection. This requires predictable transitions and known phases. In high mix assembly each fixture variant can add compliance and geometry differences. Gain sets multiply quickly and transition logic becomes brittle when parts differ from the nominal model.
The Contact Uncertainty Problem
The deeper problem is uncertain contact state. In a picking task an object may settle slightly away from its nominal position. The path follows the nominal position but actual force and torque at contact differ from the model. That difference may indicate a seated grasp or an off axis contact that fails during lift.
PID receives force feedback and responds to error but does not model correct versus incorrect contact. It minimizes position or force error against a setpoint rather than judging grasp viability. A low force error at an off center grip can still lead to a dropped part during lift.
This is where learned policies help. A policy trained on correct and incorrect grasps implicitly learns contact quality. Rather than minimizing an error term it maps observations to actions shaped by training outcomes.
Variable Payload and Inertia
High mix assembly handles parts with different mass shape and inertia. A payload adaptive PID controller needs payload mass declared or estimated before its tuning applies correctly. Operators may program each product variant. A mismatch between configuration and part creates errors.
The same arm has different dynamics with a light connector housing and a heavy motor housing. Kd suitable for the lighter part can cause oscillation on the heavier one. Kp suited to the heavier part can make the lighter part respond sluggishly.
Some robot controllers identify payload mass and center of gravity from a test motion and adjust gains. This adds calibration and assumes a rigid well characterized payload. Flexible parts loose components and liquid fill need more modeling than a rigid body payload estimator provides.
What We Use Instead
At Neuralzome we do not replace all PID control with learned policies. PID and its variants remain suitable for motion planning joint level torque control and the actuator inner loop. Those dynamics are usually well characterized and stable.
We replace the task level decision layer. It decides how to respond to the scene how to adapt to sensor reports and when to mark an attempt failed and request a retry. PID's linear setpoint model breaks down here. A policy can account for uncertainty and contact quality.
PID is not wrong. It is the wrong tool when desired behavior depends on situations that cannot be fully characterized in advance. If a mathematical model remains valid across the operating range PID or a model based variant is likely appropriate. Otherwise a learned approach can outperform further gain tuning.
A Scenario That Clarified This
During early access pilot testing in 2024 and 2025 we worked on threaded fastener insertion. The repeatable nominal operation usually worked with a PID based position controller. Cross threading occurred when slight axial misalignment made the controller drive against resistance and halt.
The halt was correct. Recovery still required a person to clear the jam back out the fastener and restart the cycle. Each incident caused 8 to 12 minutes of downtime.
We replaced the task level controller with a policy trained on normal insertion misalignment correction and deliberate cross thread recovery. It learned early force signatures and made a brief realignment move before insertion. Halt rate fell by roughly 80% in the pilot. Monitored cells needed around two hours less intervention per shift.
PID answers a specific mathematical question. When that question changes the control approach must change too.