EVA-Client: A Unified Framework for Data Collection, Inference, and Deployment of Embodied Policies on Real Robots

EVA-Client is an open-source framework for deploying, collecting data from, and evaluating manipulation policies on real robots. It sits between policy servers and physical hardware, separating robot backends, inference strategies, and transport middleware into independent layers. Its Debug, Collect, and Eval workflows support modes from open-loop simulation to continuous real-time control. Each evaluation run also records a complete rollout in a training-ready format, with exhaustive logs and a side-by-side viewer. The framework consolidates synchronous and asynchronous inference, ACT-style temporal ensembling, Real-Time Chunking, and a naive-async ablation baseline behind one configuration surface.