Publications

NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices

Filip Pawlicki, Marcel Kańduła, Marcin Pucek, Kamil Dobies

ECCV 2026 (accepted)

We present a lightweight video super-resolution system optimised for deployment on resource-constrained hardware. Our approach uses structural reparameterization to convert into standard convolutions during inference, ensuring compatibility with accelerators like TensorRT, and implicitly learns spatio-temporal alignments through progressive training rather than relying on explicit optical flow calculations. On the REDS4 benchmark, our smallest variant achieves 28.64 dB PSNR at 27.2 FPS on an NVIDIA Jetson Orin NX device, while a larger configuration reaches 29.15 dB at 19.58 FPS, demonstrating an improved balance between quality and computational efficiency for edge deployment scenarios.

Interpreting Deep Q-Networks: A Rule-Based Comparison with First-Order Logic in Wumpus World

Filip Pawlicki, Kamil Dobies, Marcin Pucek, Karol Draszawka

TASK Quarterly, Vol. 28, No. 4 (2024)  ·  Published December 2025  ·  DOI: 10.34808/tq2024/28.4/c

Deep reinforcement learning models such as Deep Q-Networks achieve strong performance across environments, but their decision-making remains largely opaque. We propose extracting symbolic rules from trained DQNs and comparing them with logic-based agents in the Wumpus World domain using decision trees and Jaccard similarity metrics. Despite comparable task performance, the two approaches employ partially overlapping but structurally distinct decision rules — highlighting the limits of behavioural equivalence as a proxy for interpretability.