Capture Regions: Transferring Insertion to Moving Fixtures by Letting Geometry Do the Contact

Robots that learn assembly in simulation now insert pegs, gears and connectors on real hardware with high reliability. They do so only when the receiving part sits still on a bench. Factory floors, vehicles, vessels and handheld fixtures move. Contact between moving parts is also the physics that simulators get most wrong: friction, stick-slip, wedging and bouncing. This proposal does not try to make the simulator more faithful at contact. It removes contact from what the learned policy is responsible for.

The idea is simple. A passive funnel is a shaped mouth around the hole that guides a slightly misplaced part inward when pushed. Such a funnel has a "capture region": the set of approach positions, angles and relative velocities from which the part slides home on its own, without jamming or bouncing out. We measure that region directly on the real hardware against a moving target. We then train the policy in simulation only to deliver the part into the region at the right moment. The policy never learns from simulated insertion contact, so simulator contact errors can no longer leak into its behaviour. The approach succeeds if a learned policy plus a funnel inserts into an oscillating fixture with a smaller sim-to-real drop than end-to-end policies trained on simulated contact.

Core Claim

The capture region of a well-chosen funnel is a quantity that transfers. Its boundary depends mainly on geometry and on a coarse friction margin, not on fine contact details. A funnel can be designed so that its region transfers from simulation to reality much better than any policy's insertion success does.

The design rule comes from the classic sliding-versus-sticking condition. Consider a peg pushed along the insertion axis against a funnel wall whose half-angle from that axis is $\phi$, with friction coefficient $\mu$. The peg slides inward rather than wedging when the tangential part of the push exceeds friction:

$$\tan\phi < \frac{1}{\mu}.$$

Wider funnels (larger $\phi$) capture larger errors but approach this jamming threshold. The friction coefficient $\mu$ is exactly the parameter that wears, contaminates and is misestimated in simulation. Our hypothesis therefore has two parts. When $\phi$ is chosen with margin below $\arctan(1/\mu_{\max})$, where $\mu_{\max}$ is a pessimistic friction bound, simulated and real capture regions should coincide closely. Near the threshold they should diverge, and that divergence should explain most of the transfer failures of end-to-end policies.

Target motion adds a third axis to the region, the relative velocity at first contact. This axis governs whether the part bounces or seats. A passive compliant wrist can widen it.

The capture region also clarifies the tension between passive alignment and safe planning around moving obstacles. Once the part enters the funnel, active correction is pointless and possibly harmful. Before that point, the arm must decide whether the predicted relative state will lie inside the region. This is a commit-or-abort decision, analogous to avoiding states from which no control can prevent failure, as formalised by Safe Motion Planning in Unknown Environments for static obstacle fields. Here it is applied to a moving goal rather than a static obstacle.

Relation to Prior Work

The closest work falls into three groups.

The first group motivates the proposal by showing what transfers on static fixtures. IndustReal transfers peg, gear and connector insertion from GPU simulation (Factory) using three components:

  • penalties that discard episodes where simulated parts pass through each other,
  • distance-field rewards,
  • an action integrator that corrects tracking error.

It still reports wedging as a residual failure. Bridging the Sim-to-Real Gap with Dynamic Compliance Tuning plans forces offline and tunes software stiffness online. Robotic Object Insertion with a Soft Wrist adds passive hardware compliance and distills a teacher policy that uses simulator-only information into a policy using force history. All three evaluate on rigid, stationary holes. All three train the policy on simulated contact. We differ on both counts: the fixture moves, and the policy is never rewarded through simulated contact.

The second group supplies the mechanism but not the method. Passive alignment works in the field without any learning. Examples include:

  • KTH's funnel fingers in the WARA challenge, which sidestepped depth-sensor failures on glass;
  • NimbRo's loosely held wrench that gravity slid onto a valve stem (MBZIRC 2017);
  • the gravity-seated surgical tools of the SASHA-OR end effector, which tolerated 225 mm of axial deviation.

These are empirical designs with static targets, no simulation model and no learned policy. We add a measured, transferable characterisation and a learner that targets it.

The third group constrains the design. Bi-Manual Manipulation and Attachment shows that heavy randomisation produces slow, conservative policies. It needed billions of steps and reached only 35% success at 1 mm precision. A further 30% of its successes came from a passive magnetic capture zone of 5–10 mm. We read this as direct evidence that a capture basin carries much of the precision burden, and as a warning against relying on randomisation alone.

Research Program

The work has four stages.

Stage 1: physical rig. We build a Franka arm running IndustRealLib with NIST Task Board 1 parts printed from IndustRealKit. The receiving board is mounted on a second arm that executes lateral sinusoids and slow tilts. The tilt mode tests gravity-dependent seating when the gravity direction moves relative to the fixture. We print a family of funnel inserts spanning half-angles from well below to near the friction threshold, and optionally mount the soft wrist design.

Stage 2: capture regions measured both ways. For each funnel we run scripted open-loop approaches on a grid of lateral offset, approach angle and contact phase of the oscillation. Phase sets the relative velocity at contact. Each trial is labelled as seated, jammed or bounced out. The same grid is run in Factory across a randomised friction range. We compare the real and simulated regions by overlap (intersection over union) as a function of $\phi$.

Stage 3: delivery policy. In simulation we train a policy whose reward is "enter the measured capture region", computed from relative pose and velocity at first contact. The policy observes the tracked fixture pose and recent motion. Episodes terminate at first contact. A commit-or-abort head predicts whether the contact state will fall inside the region and backs off when it will not.

Stage 4: deployment. On hardware the policy delivers the part and then switches to a low-gain passive mode, letting the funnel seat it.

Baselines

Naive reference. The published IndustReal policy, trained on static fixtures, is deployed unchanged against the moving board with no funnel. This measures how much the static assumption costs. For context, IndustReal reports 83–99% zero-shot real success over 600 trials on static boards. We found no reported result on Lacuna for moving fixtures.

Strongest existing method. The IndustReal pipeline is retrained in Factory with fixture motion randomised over the same amplitudes and frequencies as the rig, with the fixture's pose and velocity included in the observation, and with the same funnel inserts mounted. It is a strong competitor because it already transfers sub-millimetre contact and, given the funnel, differs from our method only in what the policy is rewarded for. It receives the same simulator, the same motion distribution, the same environment-step budget and the same hyperparameter search budget.

Soft-wrist distillation. The soft-wrist teacher-student method is retrained under target motion. It reports 100%, 95% and 80% success at 0°, +5° and −5° misalignment on a static 40 mm peg, which sets the bar for passive-compliance competitors.

Phase-wait heuristic. A scripted controller inserts only at the turning points of the oscillation, where relative velocity is near zero. It represents what good engineering achieves with no learning.

Key ablation. Using the same funnel and the same policy architecture, we train with an end-to-end simulated insertion reward instead of the capture-region reward. This isolates our main ingredient.

What counts as beating them. We claim success if our method matches or exceeds the strongest baseline's real success rate while showing a clearly smaller simulation-to-real drop, with neither result explained by slower execution. Execution time is reported alongside success.

Why not just…?

Why not just add a funnel to IndustReal and randomise fixture motion? This is the real threat, and it is the strongest baseline above. If the funnel's region transfers well, an end-to-end policy might implicitly learn to target it. Our hypothesis is that it does not, because the end-to-end policy also learns to exploit simulated sliding near the jamming threshold, which fails on real friction. The key ablation settles this. If the end-to-end reward matches ours across funnel angles, the contribution reduces to the capture-region measurement and design rule. We would report it as that.

Why not just track the target visually and servo actively? WARA teams relying on active perception struggled with transparent parts. In the bimanual attachment work, tracking error of 10 mm and 0.025 rad capped precise success at 35%. Our active-servo condition, combining vision with admittance control and no funnel, tests whether that ceiling persists under target motion.

Why not just make the arm compliant and keep a static-trained policy? Compliance absorbs force spikes but does not supply the timing that a moving target demands. The soft-wrist baseline under motion measures this directly. If it matches our method, the funnel adds nothing beyond the wrist. We expect bounce-outs at high relative velocity to separate the two.

Why not just wait for the turning point? At low frequencies this may work as well as our method, and we say so in advance. The phase-wait baseline will be swept over frequency, and we expect it to lose once the dwell time near zero velocity becomes shorter than the arm's reaction time.

Proposed Experiments and Decisive Checks

The first decisive check is the transfer curve: capture-region overlap between simulation and reality as a function of funnel half-angle. The design rule predicts high overlap well below $\arctan(1/\mu_{\max})$ and sharp loss near it. A flat curve would falsify the core claim.

The second check is the drop in success from simulation to reality, measured on NIST pegs, gears and connectors. It is reported for our method and every baseline across a grid of oscillation amplitude, oscillation frequency and tilt. Metrics are success, jam rate, bounce rate, peak contact force and cycle time.

The third check uses the commit-or-abort head. We test whether its abort decisions on hardware predict the trials that would otherwise have failed. This would show that the region serves as a usable safety criterion, not just as a training target.

The fourth check is out of distribution. We transfer to an unseen square peg, mirroring the 70% out-of-distribution result reported for the soft wrist, and to friction altered by surface contamination. Robustness should degrade only for funnels near the jamming threshold.