Piotr Sokólski

I was at @foxglove Actuate this week and it really felt like a who’s who of physical AI. I truly enjoyed meeting some of the people I had been looking up to, and definitely made new friends. Congratulations to @adrianmacneil and team for bringing together such an impressive group of people. Some takeaways inspired by this week’s conversations: Robots are having a renaissance Compared to even a year ago, I see more new companies working on different aspects of robotics, from food delivery to factory operations, to construction. And then there are robots that want to do it all. Neural Simulation vs. Rendering Engine Simulation As a founder whose company is developing a neural simulator I’m admittedly biased, but I have high conviction that combining diffusion and reconstruction is the right approach for simulation. A key advantage: Our use of Gaussian Splatting solves the domain gap problem - ie. Third Dimension’s simulator looks like you're stepping into an actual video. Not only does your simulation look like a video, but all other actors in the scene behave exactly like they did in real life - this is just not possible with classical rendering engines. On the other hand, video game-based simulators are good at providing ground truth, such as the automatic labelling of datasets, which is important for autonomy and robotics training. The downside lies in how expensive they are to build and that they just don’t look real. Both…For Now The realization for me is that there is room for both, for now. You want to get pixel perfect ground truth annotations? Tap into your classical sim engine. Want to make better use of your ever growing real world data? Neural simulation is your friend.