Kosta Derpanis (sabbatical in Zurich)#932
Some takes from the BeyondTeleop #ICRA panel, debating the usefulness of teleop data
- @JitendraMalikCV compared it to a horse controlling a car, instead recommending video + simulation pre-training + real world RL fine-tuning .
Users highlight efficient real-world RL fine-tuning as an underestimated opportunity that could advance generalist robots without teleoperation.
- @yuewang314 in other fields end-to-end learning took over model based solutions. So world models? Will we wait 50 years before simulation is solved?
-@Ed__Johns the challenge is to describe the task, hence teleop. Perhaps predict future states ?
- @JitendraMalikCV and @RobobertoMM the underestimated field is efficient real world RL fine-tuning!
At the end of the panel @kushalk_ asked the audience their opinion. More than 85% voted in favor of some teleop. We will see how it evolves next year! Thanks @altantutar
- @bgxc We show kids how to do things, we don't teleop them. Let humans show as humans, and let robots learn to be robots in simulation
- @RobobertoMM and @yuewang314 langage/vision scaled thanks to high quality data, and teleop is better than langage to describe a task
- @bgxc if we target the end game and no longer quick demos, simulation should be the focus of investment: rigid body solvers, coupling with other solvers, and assets ("an engineer problem, not a research one, just need a lot of engineers" ?)