01 · Question
What does knowing the dynamics buy?
Can an internal model reduce the corrective work left to feedback, and what happens when that internal model is wrong?
02 · Prediction
Support torque should arrive before visible error.
Gravity compensation should remove a pose-dependent load from feedback. Inverse dynamics should also use the desired acceleration. If the controller underestimates link-2 mass, it should first supply too little model torque and leave feedback with a structured residual.
03 · Experiment
Hold the physical arm fixed.
The same two-link MuJoCo plant, target, gains, and timestep ran under PD, gravity compensation, accurate computed torque, and computed torque with a controller-only link-2 mass error. The physical link remained 0.7 kg; the faulty controller assumed 0.35 kg.
Measured overlay · target reversal
The model changes where the torque comes from.
At t = 1.963 s, the desired acceleration is [−0.224, −0.288] rad/s². The gray arm is the target; the colored arms are recorded states.
Playback uses precomputed MuJoCo keyframes. The marked frame is the reversal used in the evidence.
04 · What happened
Accurate feedforward left almost nothing for feedback.
Over the final target cycle, gravity compensation reduced per-joint RMS position error [J1, J2] from [0.679127, 0.082077] rad to [0.019344, 0.004964] rad. Accurate computed torque reduced it further to [0.000414, 0.000519] rad. Its feedback-torque RMS was only [0.000536, 0.000149] N·m.
Measurement note: These vectors are per-joint RMS values, not confidence intervals. This deterministic simulation reads MuJoCo state directly and has no sensor-noise model. The matched-model position residual is also near the scale introduced by the 2 ms telemetry timing convention: the state is sampled after the simulation step while its comparison target was evaluated before it. Treat the tiny residual as a timing-limited simulation measurement, not a physical sensing-precision claim. The controller comparison is unchanged.
05 · Model update
Feedback reacts; a model can prepare.
Higher feedback gain acts after error. Gravity compensation supplies pose-dependent support before that error must appear. Computed torque also includes the torque implied by desired acceleration and motion. A wrong model creates a specific feedforward shortfall and restores feedback work; that is different from a controller that is merely poorly tuned.
Worth noting: The useful comparison was not “which trace looks cleaner.” It was whether the same physical arm moved work from error-driven feedback into a model term, and whether one wrong parameter brought that burden back.
06 · Stop boundary
The mechanism is clear enough to move on.
A later acceleration overshoot was predicted but not isolated as a separate time-resolved event. That would be a different question. This experiment established the model-based-control distinction and stops here.
Source