World Action Agent gives a visual workspace for directly piloting robots with basic tools. It centers views on contact points, previews proposed actions before execution, and supports visual correction of residual positioning errors.
The system reached 75.6% average success on LIBERO-Pro using skills evolved only from LIBERO-90, compared with 72.0% for ASPIRE. Those skills also transferred to robosuite without further learning. Separately, training Qwen3.5-9B on interaction traces increased its out-of-domain success from 1.7% to 43.3%.
The interesting claim is that changing the VLM's interaction interface, rather than training a dedicated action model, can preserve general reasoning while providing executable, revisable robot control.