CoDA: Co-Evolving Data Engines and Action Policies

CoDA uses a shared memory of policy failures and learning progress to co-evolve simulation environments and action policies. It converts vision-language critic diagnoses into generation programs to recreate and solve specific agent difficulties.