Genuine embodied agency demands that robots turn continuous real‑world experience into lasting, transferable skills. This calls for continual learning that can incorporate new capabilities without eroding prior knowledge as tasks and environments evolve. Experience replay mitigates forgetting by replaying past demonstrations, yet storing complete demonstrations becomes prohibitively expensive as the number of tasks grows. World‑action models offer a generative alternative by jointly predicting actions and future observations to reconstruct past experience. However, visually coherent rollouts may contain actions that cannot realize the predicted transitions, and adapting to new tasks can disrupt previously learned behavior. RIFAR addresses these issues by combining reliability screening with drift‑aware replay selection. It first reconstructs trajectories from compact demonstration prefixes and uses a frozen inverse‑dynamics model to assess action‑visual consistency. During training, current demonstrations are mixed with the highest‑quality screened trajectories. Afterwards, RIFAR compares action predictions before and after adaptation on identical historical inputs, and re‑selects trajectories with larger normalized drift from the same screened pool for continued training. Experiments on three LIBERO suites and real‑world robots show that RIFAR surpasses the previous state‑of‑the‑art WAM‑based generative replay. On LIBERO‑Goal, it achieves a 90.97 AUC while retaining only 320 historical time steps per task—about 4.9% of the steps kept by a 50‑demonstration replay.
Review