Robot simulators are usually judged by how far their trajectories drift from real ones, measured in meters of accumulated position error. That number can mislead. A simulated arm can follow the real arm's path closely and still get wrong how a pushed box responds to the push, and the response is what a learned controller depends on. This proposal measures simulator fidelity differently. We record the same probing motions in reality and in simulation. For each, we estimate how much the robot's recent motion helps predict the object's next move beyond what the object's own past already predicts. This quantity is transfer entropy: a directional, model-free measure of information flow between two time series. We then ask whether the simulator reproduces that pattern of influence, and whether matching it predicts which simulators produce policies that work on real hardware.
The work sits where three problems meet. Sim-to-real research lacks a fidelity score that tracks downstream transfer rather than geometric drift. Transfer entropy estimators have been validated almost entirely on synthetic dynamics, so no one knows how they behave on real contact-rich data. And evaluation that relies on simulation needs a cheap, falsifiable check against reality. Our central bet is narrow and testable. Across a population of candidate simulator configurations, the gap between simulated and real transfer entropy profiles should rank the configurations by real-world policy success better than trajectory error does. It should also do at least as well as a learned-dynamics likelihood. If it does not, we will say so, and the method's remaining value will be diagnostic.
Core Claim
Fix a probing policy and record synchronized channels: the commanded end-effector velocity $A$, the measured end-effector pose $E$, and the object pose $O$. For a source $X$ and target $Y$ at history length $k$, transfer entropy is
$$TE^{(k)}_{X\to Y} = I\big(Y_{t+1};\, X_{t-k+1:t} \,\big|\, Y_{t-k+1:t}\big),$$
the extra information the source's last $k$ steps carry about the target's next step once the target's own last $k$ steps are known. We estimate this for the links $A\to O$, $E\to O$ and $O\to E$, the last capturing the reaction force felt back through the arm. We estimate it over several lags, separately in contact and non-contact phases, and call the result the simulator's coupling profile. The fidelity gap of a simulator $s$ against reality $r$ is
$$D(s) = \sum_{\ell,k} \frac{\big|\widehat{TE}^{(k)}_{\ell}(s) - \widehat{TE}^{(k)}_{\ell}(r)\big|}{\hat\sigma^{(k)}_{\ell}},$$
where $\ell$ indexes links and $\hat\sigma$ is a bootstrap standard error pooled over probing trials. Dividing by $\hat\sigma$ keeps links that are estimated noisily from dominating the sum.
The claim has two parts. First, $D$ should correlate with real transfer success across simulator configurations more strongly than Euclidean trajectory error does. Second, adding $D$ to the objective used to tune simulator parameters should produce simulators whose trained policies transfer better than those tuned on trajectory error alone.
Why This Is Not Already Done
The robotics reality-gap work motivates the proposal. Quantifying the Reality Gap in Robotic Manipulation Tasks compared V-Rep, MuJoCo and PyBullet against 24-camera motion capture of a Kinova arm. It found free-space motion well modeled and contact badly modeled: cubes rotated 90 degrees in some engines, and PyBullet never touched the cube despite a lower wrist error (45.513 m accumulated) than V-Rep's Bullet (about 131 m). Traversing the Reality Gap via Simulator Tuning tuned up to 57 engine parameters by Differential Evolution against motion-capture trajectories and cut error by 14% to 91%. Both papers score only geometric distance, and neither checks whether the tuned simulator trains better policies.
The PyBullet example alone does not justify transfer entropy. Object displacement error would catch a missed contact just as well. We therefore treat object-trajectory error as a required baseline rather than a straw man. The subtler failure we target is a simulator whose object moves roughly the right distance but responds to the push with the wrong timing or gain. Such a simulator scores well on position error and still teaches the wrong feedback relationship.
The transfer entropy literature supplies the estimation machinery but not the setting. TENDE: Transfer Entropy Neural Diffusion Estimation estimates transfer entropy with a diffusion model, a generator that learns to denoise, whose learned gradients of log-density yield the needed divergences. It stays stable as redundant noise dimensions are added. Its only real-data test is respiration-to-heart-rate, and it assumes stationarity. Attentive Transfer Entropy weights time steps so that brief, intermittent coupling is not averaged away. Contact is exactly this kind of coupling. However, its dynamics are purely simulated neuron models, and it struggles when several drivers act at once. Causal Feature Selection via Transfer Entropy links omitted transfer entropy to prediction error for linear models.
Nonstationarity is a real constraint, since contact switches the dynamics on and off. We handle it by estimating within segmented phases rather than pretending the whole trajectory is stationary.
Domain Randomization via Entropy Maximization (DORAEMON) uses Shannon entropy over simulator parameters, not information flow over time, and supplies our transfer protocol. Finally, the synthetic-data evaluation literature frames the stakes. CausalProfiler shows that causal methods look good mainly when the benchmark's generator matches their own assumptions, and it names the missing mapping from real data back to generator settings. Is this Model for Real? compares real and simulated data by the shape of their fitted-likelihood landscapes. Our coupling profile is a time-series analogue of that "Turing test" for simulators.
Research Program
The program runs in three stages.
The first stage is simulation-to-simulation with known ground truth. We designate one physics engine, with hidden parameters, as "reality." We then use a different engine as the candidate simulator family, so the mismatch is structural and not just a matter of parameter values. We build roughly 40 candidate configurations: defaults, random draws, and configurations tuned on trajectory error. For each candidate we compute $D$ and all baseline scores from a fixed set of probing pushes. We then train a pushing policy in that candidate and measure its success in "reality." Because ground truth is cheap here, we can run many seeds and measure how $D$'s predictive power depends on the number of probing trials and on the choice of estimator (KSG nearest-neighbor, TENDE, and phase-weighted attention).
The second stage is physical. We use a box-pushing task on a real arm with motion capture, following the PandaPush setup of DORAEMON and the Kinova protocol of the tuning paper. We collect probing pushes and use them to score about ten candidate simulators, then deploy the policy trained in each candidate for 15 real trials. Mocap-derived contact intervals serve as real-world ground truth for when $A\to O$ coupling exists. This also gives a real-data validation of the transfer entropy estimators themselves, which the synthetic-only literature lacks.
The third stage is tuning. We repeat Differential Evolution tuning with the objective set to trajectory error plus $\lambda D$, holding the evaluation budget fixed, and compare policy transfer against trajectory-only tuning.
Baselines
The naive reference is the default, untuned simulator. Every candidate is also compared against random ranking, so that a correlation between a score and transfer success is read against the chance level.
The strongest existing fidelity method is Differential Evolution tuning on Euclidean trajectory error, as in Traversing the Reality Gap via Simulator Tuning, which reports 14% to 91% error reduction over defaults. We run it with the identical probing data, the same parameter subset and the same number of simulator rollouts as our tuning. Both the wrist-only and object-inclusive versions of the error are used, because the object version is the fair competitor. Beating it means that policies from $D$-augmented tuning achieve higher real success, with non-overlapping bootstrap intervals over trials. As a ranking score, it means $D$'s Spearman correlation with transfer success exceeds that of trajectory error across configurations.
The strongest transfer method is DORAEMON, which reports 60% zero-shot success on the physical PandaPush task, against 46.7% for LSDR and 0% for fixed randomization Domain Randomization via Entropy Maximization. We run DORAEMON centered on each tuned simulator with the same training compute. The question is whether a better-centered simulator improves on randomization alone.
A learned-dynamics baseline fits a one-step forward model on real probing data and scores each simulator by the likelihood this model assigns to the simulator's rollouts.
Two ablations isolate the directional ingredient. One replaces transfer entropy with its linear-Gaussian special case, Granger causality. The other replaces it with symmetric mutual information between channels, which drops the conditioning on the target's past.
For transfer entropy estimation on robotic contact data, we found no reported result on Lacuna.
Why not just…?
Why not just measure the object's trajectory error instead of the wrist's? This is the most serious alternative. It catches missed contacts and gross rotation errors, which covers the headline failures in the reality-gap study. Our claim survives only if $D$ adds predictive power once object error is already known. We test this by regressing transfer success on object error and asking whether adding $D$ improves the held-out ranking. If it does not, the claim is refuted for this task class.
Why not just fit a dynamics model on real data and score simulators by likelihood? Transfer entropy is a difference of two conditional predictabilities, so this baseline is closely related and could match $D$ as a ranking score. The advantage we expect is attribution. $D$ separates a wrong push-to-object gain from wrong free sliding of the object, telling the tuner which parameters to move. The tuning experiment tests whether that separation yields better simulators within a fixed rollout budget. If the likelihood-based tuner does equally well, the contribution reduces to interpretability.
Why not skip fidelity and randomize broadly? DORAEMON already transfers without a fidelity score. However, its backtracking causes forgetting, and fixed broad randomization failed entirely on hardware. We test whether centering randomization on a coupling-matched simulator raises success over DORAEMON centered on default or trajectory-tuned simulators.
Why not use linear Granger causality? It is cheaper. However, contact forces switch on and off abruptly, which a linear model captures poorly. The Granger ablation measures directly how much of the gain requires a nonlinear estimator.
Decisive Checks
The first decisive test is the sim-to-sim ranking experiment. If $D$ does not beat object-trajectory error and forward-model likelihood in Spearman correlation with transfer success, with confidence intervals from many seeds, we stop before hardware.
A second check is estimator adequacy. On real pushes, $\widehat{TE}_{A\to O}$ must be significant against time-shifted surrogates during mocap-labeled contact and indistinguishable from zero outside it. We will report how many probing trials this requires.
The physical stage is decisive for the tuning claim. $D$-augmented tuning must exceed trajectory-only tuning in real success at matched compute. Beyond that, we test whether pairing it with DORAEMON exceeds the published 60%.
A negative result on either hardware comparison still yields something useful: the first measurement of whether directed-information estimators, so far validated only on synthetic data, are reliable on real contact dynamics.
References
- Quantifying the Reality Gap in Robotic Manipulation Tasks
- Traversing the Reality Gap via Simulator Tuning
- TENDE: Transfer Entropy Neural Diffusion Estimation
- Attentive Transfer Entropy to Exploit Transient Emergence of Coupling Effect
- Causal Feature Selection via Transfer Entropy
- Domain Randomization via Entropy Maximization
- CausalProfiler: Generating Synthetic Benchmarks for Rigorous and Transparent Evaluation of Causal Machine Learning
- Is this Model for Real? Simulating Data to Reveal the Proximity of a Model to Reality