GapONet: Nonlinear Operator Learning for Bridging the Humanoid Sim-to-Real Gap

ICLR 2026 Conference Submission18528 Authors

Keywords: Sim-to-real gap, Nonlinear operator, Humanoid robot

Abstract:

The sim-to-real gap, arising from imperfect actuator modeling, contact dynamics, and environmental uncertainty, poses fundamental challenges for deploying simulated policies on physical robots.
In humanoids, object manipulation further amplifies this gap: end-effector payloads alter joint inertia, gravity torques, and transmission efficiency, introducing state- and payload-dependent nonlinearities. Yet existing approaches lack both systematic analysis and a generalizable representation of this payload-induced degradation.
To address this limitation, we propose GapONet, a payload-conditioned nonlinear operator that maps simulation context functions to residual actions for hardware. We then introduce a payload-aware <collect–analyze–solve> framework to learn this operator GapONet. First, we curate a sim-real paired dataset TWINS spanning multiple payloads, robots, motions, actuation rates, and simulators, comprising more than 11,298 motion sequences. Second, we perform payload-aware system identification to isolate payload-related effects and quantify their contributions, and analyze sim-to-real gaps across different simulators. Third, we train the operator GapONet to predict delta action for real-time, generalized, payload-conditioned compensation. We further introduce actuation functions and sensor predictors, which enable parallel RL training of GapONet with substantially reduced energy consumption.
While tracking unseen motions, GapONet keeps the incidence of large sim-to-real gaps below 0.09%, whereas competing methods remain near 10%. By correcting upper-body gaps, GapONet also stabilizes lower-body locomotion tracking, laying the foundation for improved performance in humanoid loco-manipulation tasks.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 18528