Reading the Reality Gap as an Equation: Per-Object Symbolic Identification for Sim-Trained Pushing Policies

Robots trained in simulation often fail on real objects because the real physics differs from the simulator's. The standard fix, domain randomization, blurs this difference by training across wide random ranges of friction, mass and other physical settings. It pays for that robustness with timid, worst-case behavior. We propose a narrower alternative. Before acting on a new object, the robot makes a few seconds of probing pushes. A pretrained transformer reads those trajectories and writes down a short closed-form equation for how this particular object slides when pushed. The simulator then practices under a tight distribution centered on that equation, instead of under generic noise.

We think this matters most in a regime that combines two known failure modes. The first is imitation learning, which here means training a policy to copy human demonstrations. Such a policy is capped by the quality and coverage of its demonstrations. The second is that policies trained on one shape break on differently shaped objects. How a pushed object slides depends on its footprint and how its weight presses on the table. Those are exactly what change when the object changes and what generic randomization treats as noise. Our bet is that this geometry-dependent part of the reality gap has compact structure that an amortized symbolic regressor can recover fast enough to use at deployment. An amortized symbolic regressor is a model trained once on millions of synthetic equations so that it proposes a formula for new data in a single forward pass. Simulated practice under the recovered law should then lift a copied policy past its demonstrations on shapes it never saw.

Core Claim

For quasi-static planar pushing of novel shapes, a transformer-based symbolic regressor can do three things from a short probe. It can recover an object-specific motion law accurate enough that simulated practice under a narrow distribution around that law transfers better than practice under broad randomization. That transfer should also be better than practice under classical parameter fitting of a standard contact model, and better than practice under a black-box neural residual model fitted to the same probe. The new element is not symbolic regression of robot physics, which Interpretable Robotic Friction Learning via Symbolic Regression already does offline for joint friction with genetic programming. It is not simulator correction either, which DexNDM does with opaque joint-wise networks. The new element is using a regressor's multiple candidate equations as a per-object randomization distribution, produced at deployment time and conditioned on geometry, and then using it to improve an imitation policy on shapes outside its demonstrations.

Concretely, let $q$ be the object's planar pose, $u$ the pusher's contact point and velocity expressed in the object frame, and $\dot q$ the object's resulting velocity. The regressor maps a probe trajectory to a small set of candidate laws:

$$\dot q = f_k(q, u; c_k), \qquad k = 1,\dots,K.$$

Here each $f_k$ is an expression skeleton from beam search, $c_k$ its fitted constants, and $w_k$ is a weight proportional to how well $f_k$ predicts held-out probe steps. Practice in simulation samples $k$ with probability $w_k$ and perturbs $c_k$ within the uncertainty of its fit. This replaces uniform ranges over friction and mass.

Research Program

The regressor follows the design of NSODE. NSODE maps sampled trajectories directly to a differential equation in one pass without finite differencing, runs in milliseconds, and is reported to be about 10x faster than ProGED. We also borrow the encoder-decoder split of NeSymReS. NeSymReS embeds an unordered point set with an order-independent encoder, decodes an expression skeleton with placeholder constants, and fits those constants afterward. Both papers supply the method. What we change is the pretraining distribution. We generate it by forward-simulating pushing on thousands of procedurally generated footprints, with randomized pressure distributions and friction laws that include anisotropic and load-dependent forms. The expressions to be predicted are drawn from a grammar of motion laws.

Two known limits of these models constrain the design. First, amortized regressors are typically limited to at most three input variables, while our inputs are closer to five. We therefore predict each output component separately and express inputs in the object frame. We treat this dimensionality limit as the main technical risk. Second, SURFACEBENCH shows that exact symbolic recovery is rare, at only 4% for LLM-based frameworks and 6% for traditional methods, and that symbolic match does not track geometric fidelity. That result keeps us from scoring laws by string equivalence. We score them only by prediction error on held-out rollouts and by downstream task success, and we treat beam diversity as useful uncertainty rather than error.

The policy side starts from Diffusion Policy, which generates multi-step action sequences by iterative denoising and reaches 95% on Push-T. We train it on demonstrations with the T-shaped block only. To go beyond the demonstrations without destabilizing the copied policy, we reuse the structure of Beyond Imitation: Q-Planning. That method keeps the copied policy frozen as a source of candidate action chunks and trains a separate value critic that picks among them. It reports that critic-guided selection raised stack-cups from 40% to 90%, where learning only from successful runs stalled at 55%. Our critic is trained only in simulation, under the per-object law distribution. Because we give it privileged state, following Asymmetric Actor Critic, no real success labels are needed. That removes the supervision bottleneck Q-planning faces on hardware.

Q-planning has a known exploration limit: if the frozen proposals never contain a valid action, no critic can rescue them. We therefore add a fallback. When the critic's best candidate scores below a threshold, a small residual policy is fine-tuned in the identified simulator. We report how often this fallback triggers.

We generate the geometric shift by morphing the T into target shapes. Greedy Shape Curriculum uses such intermediate morphs and shows that difficulty along a morph path is nonlinear. There, linear interpolation reached 50% on Cube-to-Bunny against 100% for the adaptive curriculum. We use morph distance only as an axis for measuring shift, not as a training curriculum.

To have ground truth before touching hardware, we build a "proxy reality." This is a second simulator using penetration-free incremental potential contact, following Intersection-free Robot Manipulation with Soft-Rigid Coupled Incremental Potential Contact, together with nonuniform, anisotropic friction. The training simulator uses standard rigid contact. This gives a controlled reality gap whose true law we know.

Baselines

The naive reference point is the frozen Diffusion Policy trained on T-block demonstrations and run directly on each new shape, with no probe and no practice. It measures the imitation ceiling under geometric shift. There is no reported result on Lacuna for Push-T with novel block shapes, so we measure this number ourselves.

The strongest existing competitor for transfer is broad dynamics randomization. We follow Sim-to-Real: Learning Agile Locomotion For Quadruped Robots, randomizing friction, mass, center of mass and damping over wide ranges. We use this to train the same critic, with the same critic architecture, the same number of simulation steps and the same hyperparameter search budget. No reported number on Lacuna exists for randomization-trained pushing on novel shapes. The strongest reported number anchoring the task is Diffusion Policy's 95% success on the standard Push-T benchmark, against 20% for Implicit Behavior Cloning, from Chi et al..

The strongest identification competitors see exactly the same probe data. The first is classical system identification, which fits the friction coefficient, center of mass and a pressure-distribution parameter of the standard contact model by least squares. The second is a neural residual dynamics model in the spirit of DexNDM, which fits an MLP correction to the simulator's predicted object velocity. The third is offline genetic programming with PySR, used as in the KUKA friction study, where symbolic discovery cut static-friction error to 0.033 Nm from 0.737 Nm for the symmetric physics model. PySR runs with no time limit, as an accuracy ceiling for symbolic identification. It also runs with a time budget matched to the transformer, as a latency-fair comparison.

The key ablation replaces the beam distribution with the single best equation. This isolates whether the uncertainty from multiple candidate laws, rather than the law itself, drives the gain. A second ablation shuffles probe trajectories across objects. This checks that the regressor is reading geometry-specific behavior rather than a population average.

Beating the field means two things. The method must exceed broad randomization in success rate on shapes far along the morph axis. It must also exceed classical identification and neural residuals, with non-overlapping 95% lower confidence bounds, as described below. It should approach offline PySR while taking seconds instead of minutes.

Why not just…?

Why not just fit the friction and center-of-mass parameters of the standard contact model? This is the real threat. If the proxy reality's deviations are well captured by a few scalar parameters, symbolic regression adds nothing. The KUKA friction study suggests otherwise for joint friction, where the classical symmetric model missed a four-quadrant, load-dependent structure by a factor of 22. That is joint friction, not object pushing, so the evidence is only suggestive. We settle the question by including proxy realities of graded complexity, from pure Coulomb friction with a shifted center of mass up to anisotropic, load-dependent friction. We predict that classical fitting ties the proposal on the simple end. If it also ties on the complex end, we will report that the symbolic layer is unnecessary for pushing.

Why not just run PySR on the probe data? For a single object with minutes to spare, PySR may well be as accurate or more accurate, and we report it as a ceiling. The transformer's case rests on latency and amortization. NeSymReS reports inference about three orders of magnitude faster than genetic programming at matched accuracy on AI-Feynman. That speed matters when a robot meets a stream of new objects and must identify each one in seconds. The latency-matched PySR comparison tests whether this advantage survives in our setting.

Why not just learn a neural residual model, as DexNDM does? A residual network fitted to a few seconds of probing may interpolate well but extrapolate poorly to contact configurations the probe did not visit. The critic will exploit exactly those configurations during practice. We test this directly. Probes cover only a subset of contact locations, and we measure prediction error and task success on held-out contact regions. If the neural residual matches the symbolic law there, the interpretability argument alone does not justify the method.

Why not just widen randomization? Broad randomization is known to produce conservative policies, and the file's evidence of fragility under small shifts comes from Vincent et al.. In their work, pouring success fell from 38/50 to 4/50 under a distractor, and the 95% lower bound dropped from 0.641 to 0.035. That example is a visual shift, so it does not prove the dynamics case. Our broad-randomization baseline at matched compute decides it.

Proposed Experiments

Offline, we evaluate the regressor on held-out synthetic laws and on a held-out set of unseen skeletons, analogous to NeSymReS's SOOSE split. The metric is multi-step rollout error rather than symbolic match. In proxy reality, we evaluate 20 target shapes spaced along morph distance from the T. Each method gets 200 rollouts per shape, so that confidence bounds are meaningful; Vincent et al. show that 50 trials leave wide intervals. We report success with 95% lower confidence bounds, identification latency, and fallback frequency.

On hardware, a Franka arm pushes 3D-printed morphs on a table. Each object gets a fixed probing budget, and the policy receives no real gradient updates. We run 50 trials per object and state confidence bounds explicitly rather than claiming small differences.

Three decisive checks follow. First, the per-object symbolic method should beat broad randomization on the far end of the morph axis while matching it near the T. Second, it should beat classical and neural identification on the complex proxy realities but not necessarily on the simple ones. Third, the single-equation ablation should lose to the beam distribution. If the first check fails, the proposal fails. If the second shows no separation anywhere, the conclusion is that structured randomization helps but symbolic form does not.