🎙 We held the second Lecture Club meeting for the Princeton ROB 345: Introduction to Robotics course.
Lectures 3–4 (Anirudha Majumdar): how RRT searches for a path with random points in the configuration space and how the equations of motion of a quadcopter result in a state vector. The main debate of the meeting was who makes decisions in robots: ML or algorithms.
Some thoughts from the discussion:
• Nikita - assembled an RRT sandbox with Claude for a car and saw that it is harder for it to find a path than for a drone: it won't go sideways. The main magic is that the search is not on the map, but in the configuration space: if there is no passage there, there is no path from A to B at all.
• Anatoly Stansler (product engineer) - "it blew my mind a little" that after a full graph traversal in A* and Dijkstra, RRT just places random points and strives to grow towards the unknown.
• Artemiy (warehouse robot navigation, Denmark) - roboticists do use Dijkstra and A*, and RRT works in robotic arms. Without a drone, you can take a simulator: ROS 2, Gazebo, Isaac Sim, and for depth - lectures by Cyrill Stachniss.
• Nikita - robots trained in simulation behave differently in real life: wind, resistance, and the drone constantly needs to correct itself.
• Anatoly - physics in simulation can be perfect, but if the object's center of mass is set naively, the robot will learn the wrong thing.
• Artemiy - they mainly use ML for vision, and algorithms make decisions: clients like BMW and Porsche, and a stuck robot blocks people's passage. Algorithms are more predictable.
• Konstantin Trunin - the most painful part of route finding is a good heuristic. His forecast: LLMs will suggest which direction to search and how to get out of a dead end.
• Artemiy - an algorithm gives the same answer to the same inputs. Trust in models is a matter of time, but for now, legacy and responsibility for the black box interfere.
• Nikita - from a humanoid, they expect "I bought an air conditioner, install it," and a deterministic algorithm won't cope here. Advice: interview with Physical Intelligence co-founder Sergey Levine.
📺 Meeting recording: https://youtu.be/yN0dO-vJ1dw
👍 I'm just starting to develop a YouTube channel - if the recording is helpful, like and subscribe, it really helps.
🎥 For the next meeting - Lecture 5 · Differential flatness, 67 min: https://youtu.be/64Rfzm359Xg
📆 Monday, November 2, 10:30am ET / 6:30pm MSK (the US switches to winter time, so it's an hour later in Moscow)
📍 Registration: https://luma.com/g9cn4p7h
Discuss in chat: @rvnikita_blog_chat
#robotics@rvnikita_blog #education@rvnikita_blog #ai@rvnikita_blog #princeton@rvnikita_blog
