When a robot policy trained in a simulator fails on the real robot, the engineer has to work out which of the simulator's dozens of physical settings are wrong. These include friction, damping, restitution, motor armature and camera latency. Tuning all of them together against the real trajectory usually goes badly in two ways. The search gets stuck in poor local optima. It also finds settings where two wrong values cancel each other, for example friction set too high and motor torque set too high, so the nominal trajectory looks right while each value is wrong. We propose to learn the order in which physical effects act. The method finds which parameters influence which stage of a trajectory and arranges those stages into causal tiers. It then calibrates tier by tier, fixing upstream parameters from the parts of the trajectory that downstream parameters cannot affect.
The idea combines two pieces. The first is two-stage "hybrid" causal discovery: first find which variables are connected at all, then work out the direction and layering of those connections. The second is the practical problem of closing the simulation-to-reality gap. The setting differs from textbook causal discovery in one useful way. In the simulator we can intervene on any parameter for free. On the real robot we can only watch. The first stage can therefore rely on cheap simulator interventions. The harder layering step on real data then serves a second purpose: it detects where reality's causal order disagrees with the simulator's, which is evidence of physics the simulator does not model at all.
Core Claim
The closest existing work is COMPASS. It learns a differentiable causal graph from simulator parameters and actions to the sim-versus-real trajectory discrepancy. It then restricts tuning and domain randomization (randomly varying simulator settings during training so that reality falls inside the training range) to parameters with causal effect. On a physical Kinova air-hockey task this raised policy success to 80%, against 39% for the best baseline and 0% for the untuned simulator. COMPASS has three properties that leave room for a different approach. It maps a flat set of 60 to 80 parameters to a single discrepancy. It learns the graph and the parameter values jointly by gradient descent, and its authors acknowledge this is exposed to local minima. And it assumes the discrepancy is fully explained by parameters the simulator exposes.
Our claim is that the discrepancy should not be treated as one variable. It should be split into event-segmented trajectory features: free-motion kinematics, contact onset, post-contact rebound, sensor-timed observations. The parameter-to-feature and feature-to-feature graph has a layered structure. Calibrating in that layered order should do three things that joint tuning cannot. It should shrink each optimization to a few parameters. It should break cancelling parameter pairs, because an upstream tier is pinned by features downstream parameters cannot reach. And it should expose unmodeled physics as a tier whose real-data residual no simulator parameter explains, or whose real-data causal ordering contradicts the simulator's.
Research Program
Stage one: connection pattern from simulator interventions. For each simulator parameter, we perturb it around nominal across a batch of actions and record which trajectory features move. This yields a directed bipartite graph from parameters to features. Because the evidence comes from interventions, there is no equivalence-class ambiguity on these edges. This replaces the conditional-independence tests used in the first stage of methods like HLCD and DAT-Graph. We still borrow HLCD's "lenient-then-strict" principle: keep any edge that responds under any action, then prune with a score. A parameter can matter only in rare contact regimes, and dropping its edge early is the error that hybrid methods are built to avoid.
Stage two: tiers among features. Next we order the features into hierarchical layers. We use the top-down root-finding sorts LHTS and NHTS from Hybrid Top-Down Global Causal Discovery. These identify root variables, group the rest into layers in which no later layer can cause an earlier one, and then prune spurious edges using only local neighborhoods. We run the sort twice. The first run is on simulated rollouts, which gives the simulator's causal order. The second is on real observational rollouts, where interventions are unavailable and the method's additive-noise assumptions supply the orientation. A parameter's tier is the earliest feature layer it reaches. Where the two orders disagree, for example a real-data edge from rebound angle back to mallet velocity that the simulator lacks, we flag a structural gap rather than a parametric one.
Stage three: tier-wise calibration. Let $\epsilon_k$ be the parameters assigned to tier $k$, let $f_{\le k}$ be the features in tiers 1 through $k$, and let $\hat\epsilon_{<k}$ be the already-fitted upstream parameters. We fit each tier as
$$ \hat\epsilon_k = \arg\min_{\epsilon_k} \; \mathbb{E}_{a}\, \big\| f^{\text{sim}}_{\le k}(a;\hat\epsilon_{<k},\epsilon_k) - f^{\text{real}}_{\le k}(a) \big\|^2 , $$
where $a$ ranges over executed actions. The fit uses whatever optimizer the baseline uses: gradients through a differentiable simulator, or a gradient-free method otherwise. Instead of a point estimate, we keep a distribution over $\epsilon_k$ by bootstrapping over action batches. Downstream domain randomization is then limited to this per-tier spread. This addresses COMPASS's stated limitation of point estimates.
Detecting unmodeled physics. After a tier is fitted, we test whether its real residual still depends on upstream features in a way no remaining parameter can reproduce in simulation. Such a residual signals a latent cause, which is the confounding problem that SPOT handles with ancestral graphs but only in linear-Gaussian settings. We do not claim identifiability for this detector. We treat it as a diagnostic and evaluate it empirically.
We deliberately avoid latent-hierarchy methods such as Differentiable Causal Discovery for Latent Hierarchical Causal Models. Their "pure children" requirement and their ban on observed-to-observed causation are both violated by kinematic chains. That paper constrains our design rather than supplying it.
Baselines
The naive reference is the nominal, untuned simulator. Alongside it we run uniform domain randomization over all parameters. COMPASS reports 0% real success for the nominal simulator on Mini-Air-Hockey. We will reproduce this reference in our sim-to-sim setting as the error floor.
The strongest existing method is COMPASS. It is the only published system that uses causal discovery to choose which simulator parameters to tune, and it was evaluated on the same tasks we use. Its reported figure is an 80% real-robot success rate on Mini-Air-Hockey, against 39% for the best prior baseline (the comparison included EXI-Net and NPDR). We will run it with the same rollout budget (number of real trajectories and number of simulator calls), the same optimizer, and the same hyperparameter search budget as our method. If its code is not available we will reimplement it from the paper and report that we did so. Beating it means lower parameter-recovery error and lower held-out trajectory error at equal real-data budget in the sim-to-sim setting. On any physical replication, it means real success at or above COMPASS's under matched trials.
A global black-box calibration baseline fits all parameters jointly to the full trajectory with a gradient-free optimizer at the same simulator-call budget. This tests whether any structure is needed at all.
Two ablations isolate our ingredients. Screening plus joint tuning keeps stage one (intervention-based pruning of irrelevant parameters) but tunes all surviving parameters jointly. It separates the value of tiering from the value of pruning. Hand-specified tiers orders the stages by time and physics event (free motion, then contact, then sensing) without discovery. It separates the value of discovering tiers from the value of tiering per se. A random-tier-order control checks that gains depend on the order being correct.
Why not just…?
Why not just do sensitivity screening and then tune the survivors jointly? This is the most serious threat, because in a simulator, perturbing parameters is cheap. Screening removes distractors, but it does not stop two surviving, coupled parameters from cancelling each other. Bystrova et al. (Causal Discovery from Time Series with Hybrids of Constraint-Based and Noise-Based Algorithms) show the analogous failure in causal discovery: under path cancellation, standard constraint-based methods fell to F1 around 0.4, while order-first hybrids stayed above 0.8. Our "screening plus joint tuning" ablation is exactly this alternative. If it matches full tiering on the cancelling-pair tasks, the tiering contribution is refuted, and we would report the method as a screening tool only.
Why not just order stages by time by hand, as experienced engineers do? For simple tasks this may work as well as discovery, and the hand-tier ablation will show it. Our bet is that the event order differs from the causal order in coupled tasks. In Double-Bouncing-Ball, for instance, a restitution parameter affects later free-flight segments as well as contact. Time ordering also gives no signal when reality's structure departs from the simulator's. If hand tiers match discovered tiers on all three tasks, the honest conclusion is that discovery buys automation and gap detection, not accuracy.
Why not just widen domain randomization so reality falls inside it? The real-robot COMPASS comparison already shows unstructured approaches far behind causally restricted tuning (0% nominal, 39% best baseline, 80% causally restricted). Broad randomization over 60 to 80 parameters is also known to yield over-conservative policies. We include uniform randomization as a reference, but we do not expect it to be competitive.
Proposed Experiments
Sim-to-sim benchmarks. The core study is sim-to-sim with known ground truth. We use the three COMPASS tasks: Mini-Air-Hockey with Obstacle (64 parameters), Double-Bouncing-Ball (82) and Push-I (67). A hidden "real" simulator is created by offsetting a subset of parameters. We deliberately inject cancelling pairs whose combined effect on the nominal trajectory is near zero.
Metrics. We measure four things: normalized parameter-recovery error, trajectory error on held-out actions, the fraction of cancelling pairs whose individual values are recovered within tolerance, and success of policies trained in the calibrated simulator and evaluated in the hidden one. We will also check that the discovered tiers stay stable across bootstrap resamples.
Decisive check on cancellation. The test that matters is whether full tiering beats both COMPASS and screening-plus-joint-tuning on cancelling-pair recovery at equal budget.
Unmodeled-physics test. We add a term to the hidden simulator with no exposed parameter, such as velocity-dependent drag or an actuation delay. We then measure whether the residual and order-mismatch diagnostic localizes it to the correct tier, scored by area under the ROC curve against injection location. We also check that calibration avoids distorting other parameters to absorb it, measured as the shift in recovered unrelated parameters compared with COMPASS.
Second platform. We use the REALM IsaacSim setup. Its authors identified 14 parameters, including joint friction and motor armature, and report a Pearson correlation of 0.921 between simulated and real task progression. We take their identified values as the hidden ground truth and ask whether tier-wise calibration recovers them from perturbed starts. No reported result was found on Lacuna for any automated causal calibration on REALM, so this comparison is against our own baselines only.
Physical replication. If hardware access allows, we replicate the Mini-Air-Hockey real-robot protocol and compare against COMPASS's reported 80%. Failure on any decisive check, particularly cancellation, narrows the claim to the parts that survive.
References
- What Went Wrong? Closing the Sim-to-Real Gap via Differentiable Causal Discovery
- Hybrid Local Causal Discovery
- Scalable and Flexible Causal Discovery with an Efficient Test for Adjacency
- Hybrid Top-Down Global Causal Discovery with Local Search for Linear and Nonlinear Additive Noise Models
- Scalable Differentiable Causal Discovery in the Presence of Latent Confounders with Skeleton Posterior
- Differentiable Causal Discovery for Latent Hierarchical Causal Models
- Causal Discovery from Time Series with Hybrids of Constraint-Based and Noise-Based Algorithms
- REALM: A Real-to-Sim Validated Benchmark for Generalization in Robotic Manipulation