Learning-based manipulation themes

Robotic manipulation in unstructured scenes (homes, restaurants, kitchens): perceive objects, plan motions and grasps, and control contact. MIT 6.4210/6.4212 Robotic Manipulation (Fall 2025 site; OCW Fall 2022) teaches algorithmic approaches spanning deep-learning perception, 3D geometry, planning under uncertainty, and both model-based and learning-based dynamics/control. Students build software stacks for arms in clutter, with heavy simulation (Drake) and optional hardware projects.

This is where “embodied AI” becomes concrete: not chatting about a cup—grasping it among clutter. It also shows that learning methods sit *beside* classical planning/control, not instead of them.

Perception proposes what/where objects are → planners generate collision-free and task-consistent motions → controllers execute → learning can improve any stage (vision, grasp selection, residual policies). Assignments are math + Python in cloud notebooks.

Clearing a messy kitchen counter: find the mug, plan a grasp on the handle, move without hitting the faucet, place in the sink—retry if something slips.

From the MIT course description, list the three topic buckets (perception, planning, dynamics/control). Mark which bucket a pure VLA demo is mostly selling—and which buckets still need classical engineering.

⚠️ Myth: You need prior robotics experience to start.

✓ Reality: MIT FAQ states the course assumes no prior robotics; prerequisites are basic linear algebra, probability, algorithms, and neural-net familiarity.

⚠️ Myth: Simulation means no real physics thinking.

✓ Reality: Cluttered contact tasks are chosen specifically because physics and uncertainty matter.

• MIT Manipulation Fall 2025: https://manipulation.mit.edu/Fall2025/index.html ↗
• MIT OCW 6.4210 Fall 2022: https://ocw.mit.edu/courses/6-4210-robotic-manipulation-fall-2022/ ↗
• Modern Robotics Ch. 12 grasping pointer: https://hades.mech.northwestern.edu/index.php/Modern_Robotics ↗