DrEureka: LLM automates simulation-to-real transfer without manual tuning
In one sentence NVIDIA and UT Austin present DrEureka, which uses GPT-4 to automatically generate domain randomization parameters for sim-to-real transfer. Locomotion and dexterity policies transfer zero-shot to real hardware without manual calibration.
Training a robot in simulation is much faster and safer than doing it in the real world. The problem is that robots trained in simulation often do not work well when transferred to physical hardware: the simulation is never perfect, and small differences in friction, elasticity, and sensor noise cause unexpected behaviors.
To solve this, researchers use "domain randomization": deliberately adding random variability to the simulation (different friction values, weights, noise) so the robot learns to handle uncertainty. The challenge is knowing exactly how much and what kind of variability to add — this typically requires weeks of manual work from an expert.
DrEureka uses GPT-4 to automate this process. The model analyzes the simulator code and task description, automatically generates randomization parameters, tests the results, and refines them iteratively. The robot learns to walk and perform dexterous manipulation in simulation, then transfers zero-shot to real hardware.
The practical result is that the development cycle for new robot tasks is dramatically shortened: what required weeks of manual tuning by an expert now resolves automatically in hours.
Companies
NVIDIA, UT Austin
Tools
—
Tags
Sources