Many robot assembly systems learn to put parts together by first taking them apart. Pulling a peg out of a tight hole is easy, because the part moves from a cramped configuration into open space. Pushing it in is hard, because the part must find a vanishingly small opening. So planners and learners record the disassembly motion, play it backwards, and treat the result as an assembly demonstration. This trick, called assembly-by-disassembly, now supplies reference trajectories for reinforcement learning, corrective data for imitation learning, and sequence plans for multi-part products. Every one of these uses assumes that the reversed motion is something a robot could actually execute forward in time. Friction, jamming, elastic snap-fits, and small pose errors all break that symmetry. As far as the literature in hand shows, nobody has measured how often the assumption fails, on which geometries, or whether those failures explain the assemblies on which disassembly-trained policies break down.
We propose to measure this directly. For each assembly, the reversal gap is the rate at which a compliant robot controller fails to complete the assembly when it tracks the reversed disassembly path forward under realistic perturbations. We will compute this gap for every assembly in the main public assembly-by-disassembly benchmarks, check it on hardware for a subset, and test one falsifiable claim. If reversibility breakdown is a real failure mode and not a theoretical worry, the reversal gap should predict which assemblies current disassembly-supervised policies fail on. It should do so better than simple geometric difficulty features. Fixing the gap, by replacing reversed references with forward-verified ones, should then recover a measurable part of the failures.
Core Claim
Our claim is that the reversed-trajectory assumption is the hidden reason for a specific, identifiable subset of failures in disassembly-supervised assembly, and that a cheap forward-replay test can find that subset before any policy is trained. The claim can come out false in two informative ways. If the gap does not predict policy failure beyond geometry, then closed-loop learning already absorbs reversibility errors, and the field can stop worrying about them. If the gap is large in simulation but small on hardware, or the reverse, then simulator contact models, not physics, are driving the asymmetry.
Where This Sits Relative to Prior Work
Several systems supply the method we audit. Assemble Them All plans disassembly by pushing parts with unit forces inside a physics simulator using signed distance fields (functions giving each point's distance to a surface), then reverses the plans. It reports 70.8% success on a 24-assembly rotational subset against 6.9% for geometric planners, but never executes the reversed plans forward or on a robot. State-Based Disassembly Planning solves 98.55% of 4,196 problems, evaluates only the disassembly phase, and notes that collision errors occasionally produce paths that pass through geometry. Such paths are exactly the kind our forward test should flag. AutoMate is the closest work. It turns reversed disassembly paths into dense imitation rewards for PPO, reaches at least 80% success on 80 of 100 two-part assemblies in simulation, and transfers to a Franka Panda. It attributes its 20 failures to stair-like internal geometry and grasp instability, and it does not examine whether the references themselves were forward-feasible. SRSA uses disassembly paths as features for retrieving prior skills and is restricted to linear insertions. JUICER perturbs human demonstrations backward from bottleneck insertion poses and reverses them, explicitly assuming quasistatic reversible dynamics, and reports results only in simulation. Our work differs from all of these in what it produces. It is a per-assembly, physically measured test of the shared assumption, together with a test of whether that assumption explains their failures. It is not another planner or policy.
Two papers motivate the evaluation design. FurnitureBench shows that results across labs are only comparable on shared hardware and assets, and that benchmark tests forward imitation but never disassembly-derived data. Factory supplies the fast contact simulation, with nut-and-bolt and insertion environments, that makes thousands of perturbed forward replays affordable. On Reversing Actions studies reversibility in symbolic transition systems and shows that finding reversals is computationally hard. Its notion is logical, not frictional, so it neither supports nor rules out our physical measurement. We cite it only to separate the two meanings of "reversible."
Research Program
The first component is the measurement protocol. For an assembly $a$ with reversed disassembly path $\tau_a$, we run $K$ forward rollouts in which an impedance controller tracks $\tau_a$. An impedance controller behaves like a spring pulling the part toward each waypoint, which is how deployed assembly policies interact with contact. Each rollout samples friction coefficients, part clearance (within the manufacturing range of the 3D-printed parts), initial pose error, and controller stiffness from fixed, published ranges. The reversal gap is
$$G_a = 1 - \frac{1}{K}\sum_{k=1}^{K} \mathbb{1}[\text{rollout } k \text{ reaches full insertion within tolerance}],$$
where $\mathbb{1}$ is the indicator function. Next to $G_a$ we log a failure label for every failed rollout: wedging (force rises with no progress), jamming at a waypoint, divergence caused by penetration artifacts in the source path, or missed thread or detent engagement. These labels give the taxonomy that is currently missing. We compute $G_a$ with paths generated by both the Assemble Them All planner and the AutoMate disassembly pipeline, so that planner artifacts can be separated from physical asymmetry.
The second component is hardware validation. On 20 AutoMate assemblies, 3D-printed and run on a Franka Panda with the same controller, we perform 20 forward replays each and compare the hardware $G_a$ with the simulated $G_a$. This tells us whether the simulated gap tracks a physical quantity or a simulator quirk.
The third component is a repair that serves as the causal test. We replace each reversed reference with a forward-verified one. We roll the compliant controller forward along $\tau_a$ and record the states it actually reaches. When it stalls, we apply small local search in force direction and wiggle amplitude until it progresses. The achieved forward trajectory then becomes the reference in AutoMate's trajectory-matching reward. Everything else in AutoMate's training stays fixed, so any change is attributable to the reference.
Baselines
The naive reference point is the base rate. It predicts that every assembly succeeds with probability equal to the benchmark's mean specialist success, which gives chance-level discrimination. Any predictor that cannot beat this is uninformative.
The strongest competing explanation is a geometric difficulty predictor. It is a gradient-boosted classifier on features available without any forward replay: minimum clearance, disassembly path length, number of direction changes, rotational content, contact count, and the presence of the stair-like internal features that AutoMate blames for its failures. This is the hypothesis our claim must beat, because AutoMate's own explanation of its failures is geometric. It is trained and evaluated with the same cross-validation folds over assemblies as our predictor, with equal hyperparameter search budgets. Beating it means that adding $G_a$ improves the area under the ROC curve for predicting "specialist below 80% success" by a margin outside the bootstrap confidence interval. As a further check, $G_a$ should carry significant weight in a combined model that also contains all the geometric features. No predictive accuracy for policy failure is reported anywhere in the file (no reported result found on Lacuna for per-assembly failure prediction on AutoMate), so this baseline must be run by us.
For the repair, the strongest existing method is AutoMate's own specialist training with reversed references. Its reported results are at least 80% success on 80 of 100 assemblies in simulation and 86.5% real-world specialist success across 20 assemblies (AutoMate). SRSA reports 19% higher success than AutoMate and 90% real-world success, but only on linear insertions, so we use it as a secondary baseline on the linear subset. We rerun AutoMate specialists ourselves with identical PPO budgets, seeds, and reward weights, changing only the reference trajectory. Beating it means a statistically significant gain in specialist success on the high-gap assemblies, with no loss on low-gap ones.
The key ablation isolates the forward-verification ingredient from mere reference perturbation. We train with references that receive the same magnitude of random spatial jitter as the repair introduces, but without forward verification. If jitter alone matches the repair, the benefit comes from regularization, not from feasibility.
Why not just…?
Why not just trust closed-loop reinforcement learning to absorb reversibility errors? AutoMate uses the reversed path only as a reward shape, so a policy can in principle deviate from an infeasible reference and still succeed. This is the most serious threat to the proposal, and we do not argue it away. It is exactly what the correlation test measures. If $G_a$ adds nothing over geometry in predicting specialist failure, and the forward-verified references bring no gain, we will report that the assumption is harmless for RL-based uses. The question would then shift to open-loop uses such as sequence plans and JUICER-style augmentation, where no policy corrects the reference.
Why not just use geometric difficulty features, which are cheaper? They are cheaper, and they form the main baseline above. The reason to expect them to fall short is that the same geometry can be forward-feasible or not depending on friction, clearance, and compliance. Two assemblies with equal clearance can differ in whether wedging occurs. If geometry alone explains the failures as well as $G_a$ does, our claim is refuted.
Why not just skip reversal and plan or learn forward? The file's evidence argues against this as a replacement. Forward geometric planners reach 6.9% on the rotational subset against 70.8% for disassembly-based planning (Assemble Them All), and training specialists from scratch is slower than disassembly-guided retrieval (SRSA). Our aim is to keep the disassembly trick and identify where it needs forward correction.
Why not just add more domain randomization during policy training? Domain randomization makes the policy robust to varied physics. It does not make the reference trajectory feasible. We include an arm in which AutoMate is trained with randomization widened to our perturbation ranges and the original references. If that arm closes the high-gap failures as well as the repair, the repair is unnecessary.
Proposed Experiments
The decisive simulated check runs on all 100 AutoMate assemblies in Isaac Gym using Factory contact. We compute $G_a$ with $K=256$, train specialists under matched budgets, and evaluate whether $G_a$ improves failure prediction over the geometric baseline under leave-assemblies-out cross-validation. On the 24-assembly rotational subset of Assemble Them All and the Factory nut-and-bolt environment, we report $G_a$ and the failure taxonomy for threaded and curved motions. These are the regimes SRSA excludes, and we expect the largest asymmetry there. For the open-loop case, we take FurnitureSim one_leg and compare JUICER's reversed augmentations with forward-verified augmentations. The reference point is JUICER's full pipeline at 74% on one_leg, and 71% with 10 human demonstrations (JUICER). If filtering out augmentations that fail the forward test improves success, the gap matters even when no reinforcement learning corrects it. The hardware check consists of the 20-assembly Franka replay and real-world deployment of the repaired and original specialists on the high-gap assemblies, with the same perception and grasp pipeline in both arms.
The proposal succeeds if three things hold together. First, $G_a$ measured in simulation correlates with the hardware gap. Second, it predicts specialist failure beyond geometry. Third, forward-verified references raise success on high-gap assemblies more than jitter or wider randomization do. Any one of these failing is itself a publishable bound on when time-reversed disassembly can be trusted.
References
- Assemble Them All: Physics-Based Planning for Generalizable Assembly by Disassembly
- State-Based Disassembly Planning
- AutoMate: Specialist and Generalist Assembly Policies over Diverse Geometries
- SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks
- JUICER: Data-Efficient Imitation Learning for Robotic Assembly
- FurnitureBench: Reproducible Real-World Benchmark for Long-Horizon Complex Manipulation
- Factory: Fast Contact for Robotic Assembly
- On Reversing Actions: Algorithms and Complexity