The authors test TD-MPC on 5 DM Control image-based tasks by overlaying driving videos from the Kinetics dataset as irrelevant background noise on the original observation images. Under this condition, TD-MPC's encoder, which lacks an explicit objective to filter irrelevant features, fails to retain task-relevant information and the agent fails to gain reward. The degradation is observed across all 5 tested environments and 5 random seeds, making it a reproducible failure condition triggered by the presence of visually irrelevant content in the input.
Evidence
correlational
Key metric
"bs-mpc constantly outperforms td-mpc in every environment" (Section 5.3); "with added distraction in the input image, td-mpc fails to gain rewards" (Fig. 1 caption)
Caveat
The noise is a specific type (Kinetics driving videos as background); generalisation to other forms of irrelevant input is not tested.