Track 6H · 3 stations

Robotics & sim-to-real

If the thesis is embodied.

What you will learn

A policy trained in simulation meets a real robot whose dynamics, sensors and latencies the simulator did not model. The reality gap turns a solved task into a failed one.

Domain randomisation as the standard bridge, automatic randomisation at scale in the Rubik's Cube work, and DayDreamer's world-model learning directly on hardware.

You can design a randomisation scheme for a simulator and argue about what it will and will not transfer.

Station 1 · track

Tobin et al., IROS 2017

Domain Randomization

Randomise the simulator to bridge to reality.

Automatic domain randomisation.

Resources

Station 2 · track

OpenAI, arXiv 2019

Solving Rubik's Cube with a Robot Hand

Sim-to-real plus automatic domain randomisation.

Large-scale sim-to-real.

Resources

Station 3 · track

Wu et al., CoRL 2023. Simulators: MuJoCo, DM Control, Isaac Gym, Isaac Lab

DayDreamer: World Models for Physical Robot Learning

Dreamer on real robots.

World models in the physical world.

Resources