Sim-to-Real Transfer

Sim-to-real transfer is the challenge of taking a policy trained in physics simulation and deploying it successfully on physical hardware. Training in simulation is attractive because it is fast, cheap, safe, and parallelizable — billions of experience-equivalent interactions can be generated in hours. But policies trained in simulation often fail in the real world because the simulation is an imperfect model of physical reality.

The gap between simulation and reality — the sim-to-real gap — arises from several sources:

• Contact dynamics: friction, deformation, and impact forces are difficult to simulate accurately
• Sensor noise: real sensors produce noise patterns that are hard to model precisely
• Actuator imperfections: real motors and servos have backlash, compliance, and latency that differ from simulation
• Visual appearance: real-world lighting, reflections, and surface textures differ from rendered simulation environments
• Object properties: mass, friction coefficients, and compliance vary from nominal values in ways that are hard to predict

Strategies for closing the gap

Domain randomization: Rather than trying to create a perfect simulation, randomize the parameters of an imperfect simulation — object masses, friction coefficients, visual appearances, lighting conditions — across a wide range. The policy trained across this diverse distribution learns representations robust to individual parameter values and transfers better to the real world, which is just one specific realization of the distribution.

System identification: Measure real hardware properties — motor response curves, joint compliance, sensor calibration — and calibrate the simulation to match. This reduces the gap but requires careful measurement and breaks down for properties that change with wear or environmental conditions.

Residual policy learning: Train a base policy in simulation, deploy it on real hardware, and train a residual correction policy on the real hardware to compensate for the gap. The simulation policy provides a useful initialization; the residual policy adapts to the specific deviations of the physical system.

Adaptive domain randomization: Learn to set the randomization distribution based on the distribution of real-world observations, concentrating simulation diversity where it matters most for real-world performance.

How Webbeon approaches Sim-to-Real Transfer

Object Class achieves a 14% sim-to-real gap — meaning physical performance is within 14% of simulation performance across the benchmark suite. This is achieved through:

• Aggressive domain randomization of object properties and contact parameters throughout training
• High-fidelity tactile sensing that provides a direct physical signal less dependent on visual simulation accuracy
• Residual adaptation tuned on physical hardware after simulation pre-training
• Systematic measurement of the gap at each capability tier to guide simulation improvement

Key facts

• 14% sim-to-real gap for Object Class, versus typically 30-40% for conventional systems
• Tactile sensing reduces the gap significantly because contact forces are more directly measurable than visual contact geometry
• Sim-to-real gap varies by task: basic manipulation within 4%, tool use around 12%, assembly tasks around 23%
• The gap is a research signal, not just a deployment problem — each gap component points to where simulation fidelity must improve