Humanoid-Sim2Real-DomainGapTwin

Matched-pair dataset of clean simulated vs. realistically corrupted humanoid balance-sway sensor trajectories for the IDENTICAL underlying motion: 22 sensor-suite corruption profiles, 880 matched trial pairs, for domain-randomization / sim-to-real transfer research.

About Dataset

Humanoid-Sim2Real-DomainGapTwin pairs a clean, idealized simulated humanoid single-leg-stance balance-sway trajectory with its exact corrupted "real robot" twin -- the identical underlying motion, but passed through a battery of corruption mechanisms drawn from the domain-randomization and sim-to-real literature: joint-encoder noise, quantization, and backlash; IMU bias instability, white noise, and random-walk drift (the standard inertial-sensor error decomposition); actuator/control-loop latency and jitter; and force/torque sensor gain error, fixed offset, and cross-talk. Because every "real" trial has an exact matched clean counterpart at every timestep -- not just a separately-sampled distribution with similar statistics -- this dataset directly supports measuring a domain-adaptation or sensor-fusion model's residual error against ground truth, rather than only measuring downstream task performance. 22 independently-sampled sensor-suite corruption profiles each generate 40 trials varying sway amplitude, frequency, and simulated center-of-mass height. Each trial stores both the clean ground-truth readings and the corrupted readings (14 matched channels) at 50 Hz for 2.4 seconds, plus a per-channel residual error summary. This is a fully synthetic dataset -- no real robot telemetry, proprietary, or personal data was used.

Purpose of Dataset

This Dataset Supports Domain-randomization And Sim-to-real Transfer Research For Humanoid Robotics, Specifically Enabling Direct Measurement Of A Model's Residual Error Against Exact Ground Truth Rather Than Only Comparing Aggregate Statistics Between Separately-collected Simulation And Real-world Datasets. It Targets Researchers Building Domain-adaptation Models, Sensor-fusion Correction Networks, Or Systems That Estimate The Sim-to-real Gap Itself, Across A Range Of Distinct Sensor-corruption Characteristics Rather Than One Fixed Noise Profile. The Profile-grouped Design Supports Studying Whether A Domain-gap-correction Model Trained On Some Sensor-suite Characteristics Transfers To A New, Unseen Sensor Profile.