Supplementary · OpenJar

OpenJar Evaluation Videos

Uncut rollouts from the OpenJar continual learning experiments, recorded after training on the final task.

These clips are the raw recordings behind the success rates reported on the main page. Each condition trains on the same three-task sequence with the same 5000-memory experience replay budget; only the buffer composition differs.

Condition 1 · AnchorER (Ours)

Experience replay buffer enriched with Memory Anchors extracted from the past tasks.

Episode 1

Evaluating Task 1 (Counterclockwise) after training through Task 3 (Lift).

Episode 2

Evaluating Task 2 (Clockwise) after training through Task 3 (Lift).

Condition 2 · RandER (Baseline)

Same buffer size, but memories are sampled uniformly at random from the past tasks.

Episode 1

Evaluating Task 1 (Counterclockwise) after training through Task 3 (Lift).

Episode 2

Evaluating Task 2 (Clockwise) after training through Task 3 (Lift).

Videos are hosted on Google Drive.