Most arguments about robot learning are arguments about coordinates. Model-based or learned? Sim or real? RL or imitation? Those are labels on a list, and lists don't compose — they can't tell you what a method costs, what it buys, or which methods are really neighbors.
So start with a space instead. Two axes carry almost every claim anyone makes about a robot-learning system:
- Manual Supervision — how much human effort the method consumes before it works.
- Task Complexity — how hard the problem is once that supervision is set aside.
Every method promises the same thing in these coordinates: move up without moving right. Harder tasks, less human effort. Plot the field that way and you get the map below.
The axes are projections
Neither headline axis is a real quantity. Each is a weighted blend — a one-dimensional shadow of two things that vary independently:
$$ \text{Manual Supervision} = (1-w)\cdot\text{Modeling} + w\cdot\text{Reference Data} $$
Modeling is structure written down by hand: dynamics, rewards, resets, calibration, the plan. Reference Data is behavior collected from humans: demonstrations, teleop, labels. Both cost a person's time, so both land on the same axis — and then the axis tells you nothing about which one you spent, even though they fail in opposite ways. Modeling is effort that transfers and does not scale; reference data is effort that scales and does not transfer.
Task Complexity hides a sharper split still:
$$ \text{Task Complexity} = (1-w)\cdot\underbrace{\big[(1-v)\cdot|\text{Observation}| + v\cdot|\text{Action}|\big]}_{\text{Embodiment Complexity}} + w\cdot\text{Task Dynamics} $$
Embodiment Complexity is how much robot there is — how wide the observation is to read, how many degrees of freedom there are to command. Task Dynamics is contact, underactuation, horizon, stochasticity: the part of the problem the robot cannot slow down. Sweep that slider and the map reorganizes. A generalist policy reading open-world video and a hopping robot recovering from a shove sit near each other on the projected axis and at opposite corners underneath it. The field's loudest recent progress is almost all on one child; the problems that strand a robot in the field are on the other.
Why this vantage point
This map comes out of the Robot Learning Lab at the University of Washington, from three people who have spent time on both sides of it — building learning methods, and taking robots into the field where the modeling assumptions and the demonstration distributions both run out.
Placing a system
Every dot is scored on the five atomic axes only; the headline positions are computed. Turn on place mode to drag a method to where you think it belongs (or double-click empty space to add one), then copy coordinates to export the whole set. Disagreement about a placement is the useful kind of disagreement: it is a disagreement about what a method actually costs.