Stanford OnlineStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | State Machines
Stanford Online
Reasoning with states 0:05
State machines are presented as a basic tool for giving a robot simple reasoning and decision-making. The key idea is that the output depends on the whole history of inputs, not just the current one. A state captures what matters from that history so you do not have to list every possible past sequence. A state machine is a modeling framework, not an algorithm. You still have to design the procedures that use it.
Gate control example 10:31
A gate is used to show how the idea works. The states are waiting, raising, raised, and lowering. Inputs include whether a car is at the gate, whether the bar is at the top or bottom, and whether the car has just exited. Outputs are raise, lower, or no operation. The diagram is a graph with nodes for states and labeled arcs for transitions. It also needs extra states for anomalies, because real sensors can fail or give inconsistent readings.
State reduction 18:30
Finite state machines sit low on the modeling ladder, but they are still useful. A main problem is to keep them simple and correct. One way is state minimization: you start with all states in one class, then split them when they produce different outputs or move to different states. Equivalent states are merged if they behave the same under every input.
Sequence detection example 22:22
A sequence detector for 010 or 110 can be built as a tree of states. The first digit sends you to one branch, the second digit splits again, and after the third digit the machine resets. The first version uses more states than needed because some branches carry the same meaning. After minimization, S1 and S2 are merged, and the reduced machine still recognizes 010 and 110 while rejecting other three-bit inputs.
PX4 in practice 37:30
Real flight stacks often use a finite state machine. PX4 has states such as return to land, landed, landing, and loitering. Its commander.cpp file runs a main loop at a fixed rate and checks transitions inside that loop. One example is changing the navigation mode to acro mode, which updates the vehicle status. The instructor notes that the PX4 state machine still has many open issues and active arguments in its issue tracker.
Smack and navigator 41:01
ROS can work with state machines through Smack, a framework for state and container objects. A state class defines a constructor and an execute function that returns outcomes such as success or failure. A container holds the states and their transitions. The same pattern appears in navigator.py, where a run function checks conditions like alignment or distance to the goal and then switches modes. Smack also supports sequence, cascade, parallel, and feedback style composition, plus visual display of the machine.
Smach for state machines 51:32
A planner state handles success, failure, new commands, new maps, and errors. A tracking state follows an existing plan until the end. The point is to make transitions small and clear: waiting stays waiting, a new command sends you to planning, success starts tracking, and failure sends you back to waiting. If a plan gets old, it can expire and trigger replan or stop.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
