Evo-DexBench: Benchmarking Generalization across Related Tasks for Dexterous Manipulation
Evo-DexBench systematically tests if policies can transfer knowledge from related tasks to new ones using only a few examples. It introduces the Core-Gen principle, pairing related 'Core' tasks with novel 'Gen' tasks. Experiments show that joint training of Core and Gen data significantly boosts performance, proving that functional similarity between source and target tasks is key for successful generalization.