Learning Calibratable Policies using Programmatic Style-Consistency [arxiv]

Web Demo: Basketball Player Trajectories

Instructions

In this demo, you can generate basketball player trajectories from the style consistency model and evaluate rollout style consistencies on the "Displacement," "Destination," and "Speed" labeling functions.

To generate a rollout, you can either use the ''Automatic'' simulation mode to use a starting location and style from the data distribution, or ''Manual'' mode where you must provide both the location (selected with the ''Start Location'' button) and all 3 styles. Upon pressing ''Simulate'', the desired and actual labels will be displayed below the court.

If you choose to use ''Manual'' mode, the model may make errors since the combination of starting position and joint style may be vastly different from what was encountered during training.

Labeling Functions

Destination denotes the displacement of the player's final position relative to the basketball hoop.

  • Style 1: Closeby (inside blue ellipse)
  • Style 2: Medium (inside green ellipse)
  • Style 3: Far away (outside green ellipse)

Displacement denotes the displacement of the player's final position relative to their initial position.

  • Style 1: Closeby (inside dark gray region)
  • Style 2: Medium (inside light gray region)
  • Style 3: Far away (outside gray regions)

Speed denotes the player's average speed over the course of its trajectory.

  • Style 1: Slow
  • Style 2: Medium
  • Style 3: Fast


Simulation Options