Calibrated Domain Randomization (CDR)

Updated 1 February 2026

Calibrated Domain Randomization is a methodology that adjusts simulator parameters using real-world data to bridge the gap between simulation and reality.
CDR employs techniques like moment matching, likelihood optimization, and entropy maximization to tailor training environments for improved policy performance.
Empirical results show CDR achieving up to 97% balanced accuracy in radar perception and enhanced robustness in locomotion and manipulation tasks.

Calibrated Domain Randomization (CDR) refers to a family of methodologies designed to minimize the sim-to-real gap in learning-based robotics and perception by explicitly adapting the probability distribution over randomized simulator parameters, backgrounds, or sensor statistics to match those observed in real-world datasets. Unlike classical domain randomization (DR), which samples parameters or noise from hand-tuned, typically uniform or broad distributions, CDR employs real-world calibration to focus the domain distribution and create training environments that both cover the plausible real-world variations and are solvable by the learned policy. This approach has achieved substantial gains in tasks ranging from FMCW radar perception to the control of deformable robots and robust sim-to-real transfer in complex manipulation and locomotion tasks (Trinh et al., 25 Jan 2026, Mozifian et al., 2019, Tiboni et al., 2023, Tiboni et al., 2023).

1. Foundational Principles of Calibrated Domain Randomization

The classical domain randomization strategy injects stochasticity into simulation through randomized physics parameters, sensor noise, or environmental properties. Its primary aim is to expose a learning system to broad variations, promoting robustness in deployment. However, if the randomized domain is too wide, the policy may become overly conservative or incapable of learning; if too narrow, the resultant policy may fail outside the training domain due to reality gap effects (Tiboni et al., 2023, Mozifian et al., 2019).

CDR introduces an explicit calibration procedure, using real-world measurements—either labeled or unlabeled—to infer or set the simulation parameter distribution. Calibration can be achieved by moment matching (mean, variance), likelihood optimization, or by maximizing entropy subject to empirically measured constraints. The resulting domain distribution is "locked in" to match global statistics or transition distributions observed in reality, which ensures the training distribution interpolates the real domain instead of requiring extrapolation.

2. Mathematical Frameworks and Statistical Formulation

CDR methodology is instantiated across multiple domains, with distinct parameterizations:

Noise-Floor Calibration (Radar Perception): For simulated range–Doppler (RD) maps, CDR clamps each simulated cell $S(i,j)$ with a noise sample $N_t(i,j)$ drawn from an empirically measured real-world Gaussian:

$\hat{S}(i,j) = \max\{S(i,j), N_t(i,j)\}$

where $N_t(i,j) \sim \mathcal N(m_t, s_t^2)$ , and $(m_t, s_t^2)$ are the mean and variance estimated from a calibration set of real empty-room RD frames (Trinh et al., 25 Jan 2026).

Distribution Adaptation in RL: In domains with parametric simulators, CDR is formulated as distribution learning over simulator parameters $z \in \mathcal Z$ :

$p_\theta(z) = \mathcal N(z|\mu, \Sigma)$

trained to maximize expected policy return while minimizing statistical divergence (e.g., KL) from a broad prior:

$\mathcal L(\theta) = \mathbb{E}_{z \sim p_\theta}[J_{M_z}(\pi)] - \alpha D_{KL}(p(z) \| p_\theta(z))$

Iterative updates regularize $p_\theta$ to keep it close to the prior while focusing mass on "solvable" parts of the domain (Mozifian et al., 2019).

Entropy Maximization Under Success Constraint: The DORAEMON algorithm casts CDR as a constrained entropy-maximization problem:

$\phi^* = \arg\max_\phi H(\nu_\phi)\quad \text{s.t.}\quad G(\theta, \phi) \ge \alpha, \quad D_{KL}(\nu_\phi \| \nu_{\phi_{\text{prev}}}) \le \epsilon$

where $H$ is the entropy of the parameter distribution, $G$ is in-distribution policy success, and $\epsilon$ is a KL trust region (Tiboni et al., 2023).

3. Algorithmic Pipelines

Common CDR pipelines share the following stages:

Stage	Description (Example from (Trinh et al., 25 Jan 2026))	Input/Output
Sim-data Generation	Physics-informed geometric simulation, e.g., corridor geometry for radar perception	Simulated data frames
Real Calibration	Collect real data (e.g., empty room radar, baseline trajectories), extract statistics	Empirical moments / likelihoods / success rates
Calibrated Injection	Inject calibrated statistics (Gaussian noise, likelihood-weighted parameters)	Simulated frames with real-matched statistics
Policy/Model Training	Train perception or control network solely on calibrated simulated data	Policy or classifier
Real-world Evaluation	Evaluate on held-out real data, compute accuracy or return metrics	Quantitative sim-to-real performance

For soft robots, CDR is performed by inferring a multivariate Gaussian over dynamics parameters via likelihood maximization on real-world transition batches, followed by DR policy training using the calibrated distribution (Tiboni et al., 2023). In DORAEMON, each iteration alternates RL policy updates with parameter distribution expansion, until the highest-entropy distribution is achieved given a minimum success rate (Tiboni et al., 2023).

4. Quantitative Performance and Comparative Analysis

CDR consistently outperforms standard DR and simulation-only training:

Radar Occupancy Detection (Trinh et al., 25 Jan 2026):
- CDR yields 97% balanced accuracy (binary occupancy), compared to 83% for uncalibrated DR and 50% (chance) for pure ray tracing.
- People counting achieves 72% balanced accuracy (CDR), 61% (DR), 33% (RT).
- Reductions in false positives/negatives are substantiated by confusion matrices.
Locomotion Policies (Hopper, Half-Cheetah) (Mozifian et al., 2019):
- LSDR (learned DR, a CDR instantiation) converges domain distributions to bands where RL policies are reliably solvable and retains high coverage, outperforming both fixed uniform DR and exhaustive grid training.
Robotic Manipulation (PandaPush) (Tiboni et al., 2023):
- DORAEMON achieves 66.6% sim2sim and 60% sim2real success, with mean error 2.68 cm, while all baselines either fail to transfer (≤46.7% success) or require substantially larger final push errors.
Soft Robot Control (Tiboni et al., 2023):
- CDR-DR policies match oracle performance in endpoint error, remain robust under parameter misspecification (Young’s modulus fixed 80% off), and permit use of drastically simplified models for significant speedup without loss in transfer accuracy.

5. Assumptions, Limitations, and Open Challenges

CDR carries several assumptions:

Calibration steps typically assume that real-world noise or parameter statistics can be well-approximated by unimodal, independent distributions (Gaussian or Beta families).
Only first and second moments are matched; structured clutter, correlated noise, or higher-order statistics are generally not addressed (Trinh et al., 25 Jan 2026).
Real calibration requires a preparatory data collection pass, e.g., empty room radar frames or batch trajectories under known controllers.
Offline CDR inference (likelihood-based parameter fitting) incurs significant compute load for repeated simulation but is often amortized over short data batches (Tiboni et al., 2023).
Success criteria (for entropy-maximizing CDR) may be heuristic or empirically tuned, e.g., fixed thresholds for task completion in RL (Tiboni et al., 2023).
Coverage guarantees and monotonic improvement of policy robustness under learned DR distributions remain subject to ongoing research (Mozifian et al., 2019, Tiboni et al., 2023).

A plausible implication is that in highly non-Gaussian or multimodal domains, extending CDR to richer distribution families (mixtures, flows) may be necessary for complete coverage.

6. Best Practices, Extensions, and Generalization

Adoption of CDR involves the following practitioner steps:

Identify minimal, physically meaningful dynamics parameters to randomize.
Choose initial, broad search intervals and priors; calibrate using short real-data sequences (Tiboni et al., 2023).
Employ Bayesian, evolutionary, or gradient-based optimizers for offline calibration.
During policy training, sample from the calibrated distribution only; retrain if domain shifts are detected.
For entropy-constrained approaches, use trust-regions to maintain exploration while guaranteeing success.
In domains where new real data become available over time (e.g., wear, aging), re-run CDR to re-calibrate without manual intervention.

Extensions include:

Joint optimization of policy and simulator distribution in a unified Lagrangian framework (DORAEMON) (Tiboni et al., 2023).
Hybrid CDR methods that mix simulated and real data for online domain refinement.
Direct application to discrete parameter spaces and high-dimensional environments.

7. Contextual Significance: Impact Across Robotics and Perception

The emergence of CDR formalizes the empirical intuition that interpolation between calibrated real-domain statistics and simulated platforms is preferable to extrapolation over arbitrarily wide parameter spaces. CDR has unified sim-to-real transfer in radar, soft robotics, and manipulation by automating environment distribution inference, reducing false transfer failures, and dramatically improving data and computational efficiency. It provides algorithmic assurance that the training ensemble covers the true deployment domain, which is essential for both robust perception and control policy generalization (Trinh et al., 25 Jan 2026, Mozifian et al., 2019, Tiboni et al., 2023, Tiboni et al., 2023).