From goal to action
A household or industrial task can require hierarchical decisions: interpret the goal, select a sequence of sub-tasks, choose grasps or trajectories, then execute while reacting to feedback.
Research / Action
Planning & Control
Planning determines what a robot should do; control turns those decisions into stable physical behavior. Embodied AI increasingly treats the two as connected layers rather than isolated modules.
Why control matters
Real hardware has dynamics, latency, friction, collisions, actuator limits and uncertainty. A successful high-level plan can still fail if low-level execution cannot maintain balance, contact or precision.
Learning-based control
Imitation learning, reinforcement learning and vision-language-action policies can learn parts of the mapping from observations and goals to actions. Practical systems still need interfaces to sensors, actuators, safety constraints and task execution.
Research questions
- How much planning should be explicit versus learned?
- How can policies recover from execution errors?
- How should whole-body constraints shape manipulation?
- How can learned controllers generalize across environments and embodiments?