A new paper on arXiv introduces SHRED, a self-play method for modeling interactions among heterogeneous road users. The authors note that realistic traffic simulation and autonomous planning both depend on capturing how different agents—such as cars, bicycles, and pedestrians—interact, despite their fundamentally different dynamics and constraints.
SHRED is designed to learn these interactions without expert demonstrations, which are often expensive or difficult to obtain. Instead, the method leverages self-play, allowing agents to generate their own training signal through interaction.
The paper's abstract is brief, but it clearly positions SHRED as an alternative to existing approaches that predominantly learn from demonstration data. The full details of the method and its evaluation are not included in the abstract, so the reported results and comparisons remain to be seen.