Continual Lifelong Learning in Robotics: Open Challenge

Comparing 6 continual learning strategies for robotics: accuracy, forgetting, forward transfer, and lifelong learning score.

0.975

Best Accuracy (Progressive)

0.025

Lowest Forgetting

0.850

Best LLS

6

Strategies Compared

0.382

Best FWT

Average Accuracy by Strategy

Forgetting vs. Forward Transfer

Lifelong Learning Score (LLS)

Backward Transfer (BWT)

Full Strategy Comparison

Strategy	Avg Accuracy	BWT	FWT	Forgetting	LLS
Progressive Nets	0.9753	-0.0175	0.3818	0.0251	0.8499
Adapter Routing	0.9725	-0.0166	0.3778	0.0250	0.8487
Experience Replay	0.9253	-0.0773	0.3794	0.0794	0.8075
PackNet	0.8818	-0.1280	0.3750	0.1290	0.7700
EWC	0.8541	-0.1669	0.3776	0.1669	0.7438
Naive Finetune	0.7531	-0.2980	0.3820	0.2980	0.6511

Key Insights

Aspect	Finding
Top Performer	Progressive Networks achieves highest accuracy (0.975) and lowest forgetting (0.025)
Architecture Methods	Progressive Nets and Adapter Routing dominate over regularization approaches
EWC vs PackNet	PackNet (pruning) outperforms EWC (regularization) by 3.2% accuracy
Forward Transfer	All methods achieve similar FWT (~0.38), suggesting FWT is strategy-independent
Accuracy-Forgetting Tradeoff	Strong inverse correlation: lower forgetting yields higher accuracy