Publikation
Do Object Channels Improve Robustness in Deep Reinforcement Learning?
Jannis Blüml; Cedric Derstroff; Bjarne Gregori; Elisabeth Dillies; Quentin Delfosse; Kristian Kersting
In: Trans. Mach. Learn. Res. (Hrsg.). Transactions on Machine Learning Research (TMLR), Vol. 2026, Pages 1-40, TMLR, 2026.
Zusammenfassung
Pixel-based reinforcement learning agents often exploit spurious visual correlations, leading
to brittle policies that fail under minor visual perturbations. We systematically investigate
spatial grounded semantic channel representations, often called Feature Maps, Planes, or
Object Channels, as a representation design principle for reducing shortcut learning. Object
channels map detected entities into binary tensors aligned with the original coordinate frame,
preserving compatibility with standard RL backbones without architectural modifications.
Specifically, through systematic evaluation in Atari environments under controlled pertur-
bations, we demonstrate that such channel representations substantially improve zero-shot
robustness to distribution shifts while maintaining competitive in-distribution performance.
We analyze the abstraction–fidelity trade-off and show that combining object channels with
raw pixels improves robustness and sample efficiency compared to pure pixel-based ap-
proaches. The experimental results indicate that spatially grounded object-based encodings
offer a practical mechanism for bridging pixel- and object-centric RL.