Prof. Anima Anandkumar

I am reviewing some CoRL (Conf. on Robot learning) and Humanoids papers (6 papers total). None of them are reporting tests with a real robot (only simulation) => When did the robotics community stop caring about actual robots? these papers should be submitted to SIGGRAPH!

I disagree. There are now realistic large scale simulation tools that enable us to experiment in a completely different way compared to getting measurements with a real robot. algorithmic questions on how to design robust sample-efficient interactive learning is best studied here

Siggraph would be a terrible fit for such papers. The emphasis there is on visual appeal not on learning algorithms or physically valid simulations. Robotics community needs to treat simulations as a core topic

I would love to have more papers about better simulations of robots, especially if there is some data to compare them to the reality!

It has been shown that for many grasping tasks simulations have gotten very realistic. Dynamics is harder and we will probably never get it completely right.

Yes, simulating a position-controlled arm is working well, but even in this case there is often a reality gap (ask @sim2realAIorg !) because of the context models. Things are harder in torque control and harder for dynamics like (walking).

I agree sim2real will be present but robotics research is much more than that. Sample efficient reinforcement learning, unsupervised learning, safety and stability issues are all important. They should be encouraged to thrive in same community