The Sequence Scope: ML, Physics and Robotics

Weekly newsletter with over 100,000 subscribers that discusses impactful ML research papers, cool tech releases, the money in AI, and real-life implementations.

TheSequence | Ksenia Semenova | Substack

📝 Editorial: ML, Physics and Robotics

Robotics is one of the quintessential use cases for machine learning (ML) but also one of the most difficult ones to implement. Robotic applications need to combine hardware and ML models in very complex ways and are subjected to interactions with real-world environments, which are very hard to model. Object interactions, texture, environmental dynamics, advanced geometry are common elements that need to be present in physic simulation models. Not surprisingly, physics simulation is one of the hardest things to get right in ML solutions. The absence of robust simulators remains one of the biggest obstacles to making robotics research mainstream. Many of the physic simulation stacks remain closed-source and based on proprietary solutions. However, the ML industry has been making steady progress in this area.

A few days ago, DeepMind announced that it was acquiring and open-sourcing MuJoCo ( Multi-Joint dynamics with Contact), one of the most popular frameworks for robotics research. MuJoCo is a general-purpose physics engine that enables robust simulation of articulated physical structures and environments. Open-sourcing MuJoCo and combining with DeepMind’s advanced ML models in areas such as reinforcement learning could really accelerate robotics research and applying ML in physics in general.

🗓 Next week in TheSequence Edge:

Edge#135: we discuss Self-Supervised Learning for Computer Vision; we explore SEER, one of the most powerful SSL models for computer vision ever built; we cover Hugging Face library for computer vision, including such models as Vision Transformer, VisualBERT, and DeIT.