Fig. R1. Success rates (%) on BabyAI GoToSeq task with varying numbers of demonstrations. Fig. R2. We visualize two LOReL episodes with aligned keyframes and the discrete skill index selected at each timestep. Top: instruction “turn faucet right and close drawer.” The skill sequence [16, 16, 4, 9, 1, 1, 6] tracks sub-stages. Bottom: “open drawer and move black mug right” with [0, 18, 12]. (a) LISA (b) DASL Fig. R3. LISA and DASL codebooks. Heatmaps visualize code correlation, option frequency, and word frequency for each method. Fig. R4. Real-world scenario. The assets and environment configured for the real-world experiments. Fig. R5. Open drawer task. Snapshots at different timesteps.