A new benchmark called Evo-DexBench, described in a recent arXiv preprint, targets a core challenge in dexterous manipulation: getting policies to transfer knowledge from related tasks to new ones with only a handful of demonstrations. The authors argue that existing benchmarks are unrealistic because they provide too many target demonstrations, making the evaluation of few-shot generalization difficult.

Evo-DexBench is designed to measure how well a policy can adapt across related manipulation tasks, rather than just mastering a single task with abundant data. The abstract notes that current benchmarks "also leave…" but the sentence is cut off in the source, so the full list of limitations is not available.

According to the abstract, the benchmark is intended to push toward general-purpose dexterous manipulation, where a robot can quickly learn new variations of tasks it has some experience with. The specific tasks, evaluation protocols, and results are not detailed in the provided text, so this summary is limited to the stated motivation and focus.