Real-IKEA : Simulating What Robots Will Really See and Touch

ICLR 2026 Conference Submission1602 Authors

Keywords: Contact-rich Robotic Manipulation, Articulated Object, Simulation Fidelity, Sim-to-Real Gap

Abstract: Robotic manipulation has greatly benefited from simulated data, yet in contact-rich tasks policies often fail to transfer. We trace this sim-to-real gap to three sources: object assets, physical realism and visual fidelity. We emphasize accuracy along all three axes—precise meshes and collisions, calibrated friction and hinge resistance, and visually realistic observations—and present Real-IKEA, a dataset and simulation framework designed with accuracy as a first-class goal. At scale, Real-IKEA provides 1,079 articulated asset configurations, created by combining real IKEA furniture bases with a curated library of 83 authentic IKEA handles and knobs. For contact-geometry accuracy, we introduce a bidirectional surface-deviation metric ($E_{Q\to P}$, $E_{P\to Q}$) that quantifies collision meshes against the visual mesh. For dynamics accuracy, we establish resistance-calibrated benchmarks that vary damping and friction. To narrow the vision gap, we pair real-time teleoperation with offline high-fidelity re-rendering and quantify alignment via FID/EMD across multiple encoders. Extensive comparisons show that Real-IKEA yields more realistic asset structure, more accurate physical interactions, and visuals more closely aligned with real data, enabling policies to exploit geometry and torque rather than rely on friction-only pulling. This accuracy-centric design, coupled with large scale, enables the scalable collection of reliable manipulation data and more robust sim-to-real transfer.

Primary Area: datasets and benchmarks

Submission Number: 1602