Spending Simulation Where the Flange Saturates: Pareto-Targeted Synthetic Data for Macro–Micro Finishing

Robots that grind, polish, or deburr surfaces often split the job between two actuators. A heavy industrial arm follows the part slowly and coarsely. A small, fast actuator at the wrist, such as a pneumatic "active contact flange," holds the contact force steady against small bumps. The operator then faces a real trade-off. Moving the arm faster shortens cycle time, but the small actuator eventually runs out of stroke or bandwidth, and at that point force spikes or contact is lost. Learning a policy that picks feed speed and force setpoints for a requested trade-off requires many contact simulations. Most of that compute is wasted on easy segments where the small actuator absorbs everything and nothing interesting happens.

This note proposes using the decoupling itself as a guide for where to spend simulation. While the wrist actuator is inside its stroke and bandwidth limits, contact force is nearly independent of arm error, so a cheap linear model of the actuator predicts the outcome well. Expensive, fine-timestep contact simulation is only needed near predicted saturation. Those saturation events are also what bend the speed-versus-quality trade-off curve. A data generator that runs the cheap model everywhere and reserves high-fidelity simulation for low-margin segments should therefore recover the trade-off curve at a fraction of the compute. The resulting data would train a small, preference-conditioned scheduler for the slow arm layer, leaving the fast force loop untouched.

Core Claim

In a macro–micro decoupled finishing system, two things coincide in the same small set of trajectory segments: the error of a cheap decoupled surrogate model, and the shape of the Pareto front over cycle time, force quality, and actuator saturation. The Pareto front is the set of behaviors where no objective can improve without another getting worse. Because these two things coincide, saturation-margin–guided allocation of high-fidelity simulation yields the same front, measured by hypervolume, as uniform high-fidelity generation, at several times lower compute. A student policy distilled from this data and conditioned on the user's preference generalizes to unseen part geometries better than per-part tuning.

The claim is falsifiable at its premise. If surrogate error is not concentrated at low saturation margin, the method has no reason to work, and the first experiment checks exactly this.

Positioning

Three lines of prior work motivate the gap. Assessment of force control for surface finishing shows that a dedicated active contact flange settles in 0.22 s against 1.00 s for the UR10e's internal controller, and holds about 2.7 N mean absolute error against about 4.6 N at 0.9 m/s. Its gains and speeds are fixed, so it never maps how performance degrades as speed rises. The active tool holder and active worktable work, such as Robotized bead grinding task using active tool holder and An active worktable for force control of robot manipulators, establishes the hardware split we assume. It is likewise tuned by hand for one operating point.

CHEQ-ing the Box learns impedance and poses for curved-surface polishing with a single scalar reward at 20 Hz. It needs about 250,000 physical steps (about 8 hours) for one workpiece. We differ in three ways: we keep a fixed fast force loop, we learn only the slow layer, we target a whole trade-off curve rather than one scalar optimum, and we train from simulation.

Two works supply design constraints. TacDiffusion shows that a slow learned policy (50–150 Hz) feeding a 1 kHz loop needs a dynamic filter; removing it cost 9.15% success. This is why our learned component only sets slow references for the flange and never commands force directly. Interactive Force-Impedance Control shows that contact loss and energy injection threaten stability. We therefore keep a certified inner loop and treat saturation and contact loss as explicit objectives rather than rare failures.

On the data side, MimicGen scales about 200 demonstrations to more than 50,000 by pose transforms plus simulate-and-reject. Its generation is kinematic and samples scenes uniformly, so its rejection compute is spent blindly. Toward synthetic data generation for robotic tactile manipulations reports finite-element contact simulation to be the bottleneck. DISCO shows that 75,000 verified samples beat more than 1,000,000 unverified ones, which supports targeted verification over volume. Our contribution is a physics-derived rule for which samples deserve expensive verification.

For the multi-objective learner we reuse ideas rather than invent them. Preference-conditioned single networks follow Basaklar et al.. MORL-FB is our strongest learned competitor, having reported Pareto coverage with about 1/21 of the samples of per-preference training on MO-Gymnasium. The warnings that weighted sums miss non-convex front regions (Cheng et al., Qiuyi Zhang) matter here because saturation produces a sharp knee in the front.

Research Program

The system. The system is a position-controlled arm carrying an active contact flange that regulates normal force at about 1 kHz with its own fixed controller. The learned slow layer runs at 50 Hz. It outputs three things per segment: tangential feed speed, the flange force setpoint, and a normal offset of the arm that re-centers the flange stroke ahead of upcoming geometry. Its inputs are a short preview of surface height and curvature along the path, the current flange stroke, and a preference vector over three objectives:

  • cycle time;
  • force quality, defined as force mean absolute error plus a peak-force penalty;
  • saturation time, the fraction of time the flange is at a stroke or bandwidth limit, with contact loss counted as saturation.

Two simulators. The cheap surrogate convolves the geometry error seen by the flange with an identified linear model of the flange (pneumatic dynamics plus tool and contact stiffness). It returns force and stroke traces in milliseconds per segment. The expensive simulator runs fine-timestep rigid-compliant contact with friction, the full flange dynamics, and arm tracking error at 1 kHz. It also includes a second-tier compliant-tool model on a subset of segments, in the spirit of the finite-element pipelines whose cost motivates this work.

For each segment, the surrogate predicts a saturation margin

$$m = 1 - \max_t \frac{|z(t) - z_0|}{z_{\max}},$$

where $z(t)$ is predicted flange stroke, $z_0$ its center, and $z_{\max}$ its half-range. A margin near zero or negative means the flange is predicted to bottom out. The expensive simulator runs with probability $p(m) = \min(1, \exp(-m/\tau))$, where the temperature $\tau$ sets the compute budget. Its result also corrects the surrogate's label for that segment.

Data and training. The generator samples procedurally varied surfaces (ramps, steps, weld-bead profiles, curved bridges) and preference vectors. For each pair, a sampling-based optimizer (cross-entropy method) chooses segment-wise speed, setpoint, and offset under the multi-fidelity simulator, producing teacher schedules. A small preference-conditioned network is then distilled from these schedules. As an alternative learner, MORL-FB is trained on the same multi-fidelity environment, which separates the effect of the data generator from the effect of the learner.

Analyses. Beyond hypervolume, the program measures three things. First, surrogate error as a function of $m$, which tests the core premise. Second, how many expensive rollouts each method spends per point of hypervolume gained. Third, where on the front the knee caused by saturation sits, and whether weighted-sum methods find points beyond it.

Baselines

Naive reference. The flange runs with a fixed setpoint and the arm runs at a fixed, conservatively hand-chosen feed speed, which is how the flange was evaluated in Assessment of force control for surface finishing. That paper reports about 2.7 N mean absolute error at 0.9 m/s and 0.22 s settling on a ramp. These are the hardware numbers our scheduler must not degrade at matched speed. Beating it means strictly lower cycle time at equal or lower force error on held-out parts. We will sweep the fixed speed to draw this baseline's own front, so the comparison is front against front rather than front against a single point.

Per-part Bayesian optimization. This baseline searches a constant speed, setpoint, and offset for each part, using the expensive simulator at the same total compute. It is the most dangerous competitor on parts it was tuned for. Beating it means higher hypervolume on held-out geometries and under preference changes it was not tuned for. On seen parts we expect it may tie or win, and we will report that.

Scalarized RL. Soft Actor-Critic is trained on a weighted sum of the three objectives with a CHEQ-style reward on the same slow-layer action space. The compute budget is split across five weight settings, all using the same multi-fidelity simulator. This isolates whether a preference-conditioned front is needed at all.

MORL-FB. The strongest existing multi-objective learner (MORL-FB) is run twice: once on uniform high-fidelity data and once on our generator. No absolute hypervolume values are reported for it on Lacuna, so we will reproduce it ourselves with equal environment steps and tuning trials.

Uniform generation with rejection. This baseline samples surfaces uniformly, simulates everything at high fidelity, and filters failures, as in MimicGen. It is matched in total CPU-hours. No reported task-success result was found on Lacuna for MimicGen to anchor against, so it is reproduced here.

Ablations. Two ablations isolate the key ingredient. Random allocation spends the same fraction of expensive rollouts but chooses segments at random rather than by margin. Surrogate-only uses no expensive simulation. If random allocation matches margin allocation, the central claim is wrong.

All methods share the same surfaces, held-out split, preference grid for evaluation, and 1 kHz flange controller.

Why not just…?

Why not just tune feed speed and setpoint per part with Bayesian optimization? For a single high-volume part this may be the right answer, and the per-part baseline is included to show where that is so. The weakness is that each new geometry or each change in operator priority, for example a rush order favoring cycle time, needs a fresh search. Constant settings also cannot re-center the flange ahead of a step. The held-out-geometry and preference-shift tests settle whether the learned scheduler is worth its cost. If it does not beat per-part optimization on held-out parts at matched compute, we will report that the program fails.

Why not just use a weighted-sum reward with a few weight settings? Saturation creates an abrupt knee: past some speed, force error jumps rather than rising smoothly. Cheng et al. and Qiuyi Zhang argue that linear scalarization misses non-convex regions of exactly this kind. The scalarized-RL baseline tests whether this matters in practice. If its five runs cover the front as well as our method, the preference-conditioned student is unnecessary, and we would keep only the data generator.

Why not just spend more compute on uniform high-fidelity simulation? This is the honest competitor to the core claim. The compute-to-hypervolume curve answers it directly. If the uniform curve reaches the same hypervolume within a small constant factor of compute, margin-guided allocation is not worth the added machinery. DISCO's finding that verified data beats sheer volume (DISCO) suggests targeting matters, but it comes from text and is not proof for contact physics.

Why not just learn on the hardware with a safety wrapper? CHEQ needed about 250,000 steps and 8 hours for one workpiece (CHEQ-ing the Box). Doing that for each part, preference, and tool-wear state wears tooling and occupies the line. Hardware is reserved here for validating the transferred policy.

Proposed Experiments

Premise check (simulation). On 500 held-out segments, we compare surrogate and high-fidelity force and stroke traces and plot error against margin $m$. The decisive outcome is that the large majority of surrogate force error falls in a low-margin minority of segments. If error is spread evenly, we stop and report a negative result.

Main simulation study. We generate data under budgets spanning roughly an order of magnitude in CPU-hours and plot hypervolume over the three objectives against compute for every baseline. Evaluation uses held-out surfaces and a fixed preference grid, with the reference point set by the naive fixed-speed policy. Success means reaching uniform generation's final hypervolume at several times less compute, and beating the random-allocation ablation at every budget.

Hardware transfer. On a UR-class arm with an active contact flange and an external force sensor, we reproduce the ramp, straight-line, and impact profiles of Assessment of force control for surface finishing, plus weld-bead and curved-bridge parts not seen in training. For three preference settings we record:

  • cycle time;
  • force mean absolute error;
  • peak force;
  • settling time;
  • flange saturation fraction.

Success means that at the quality-favoring preference the scheduler matches the flange-only baseline's 2.7 N error at a higher average feed. At the speed-favoring preference, it should cut cycle time without the force spikes the fixed-speed baseline shows when pushed to the same speed.