EVA-Client: New open-source framework unifies robot policy deployment and data collection

EVA-Client is a new open-source framework designed to unify the entire real-robot policy iteration loop — data collection, inference, and deployment — all within a single codebase. Developed by Heqing Yang and colleagues, the framework sits between a policy server and physical hardware, providing a clean abstraction layer. Its key architectural innovation is a component-decoupled design where robot backends, inference strategies, and transport middlewares form an orthogonal grid. This means adding a new robot arm (e.g., a Franka Emika Panda) or switching an inference strategy (e.g., from synchronous to asynchronous execution) only touches its own layer, drastically reducing integration complexity.

The framework supports three inspectable workflows — Debug, Collect, and Eval — that span from open-loop simulation to continuous real-time control. A standout feature: every evaluation run automatically doubles as a data collection session, recording full rollouts in a training-ready format, complete with exhaustive logs and a side-by-side comparison viewer. This turns each test into a reusable asset for the next round of model training. EVA-Client also consolidates major real-time inference strategies — including ACT-style temporal ensembling, Real-Time Chunking, and a naive-async baseline — behind a single configuration surface, making it easy for researchers to benchmark and iterate.