Fig. 4: Evaluation of replay’s contribution in knowledge recall.

a, The average success rates from five runs of the LEGION framework compared with three replay-based lifelong learning methods: perfect memory, reservoir and A-GEM. The figure shows that LEGION consistently outperforms these methods, demonstrating a steady increase in success rate throughout the task sequence. b, Evolution of the ‘push’ task data ratio within the training batch. While the batch size remains constant, the data ratio for the ‘push’ task gradually decreases from an initial maximum of 50% to 10% after learning 10 tasks. c–g, Few-shot knowledge-recall performance on reach (c), push (d), faucet open (e), button press (f) and window close (g). The agent is trained sequentially on five selected tasks over three repeated loops, with buffer capacity limited to three tasks at a time. This configuration forces the agent to pause on each base task for 1 million steps without replay. For a and c–g, the data are calculated based on at least five trials, presented as mean ± standard deviation (μ ± σ).