Noietch/EVA-CLIENT

What it solves

EVA-Client provides a unified software stack to bridge the gap between AI policy development and real-world robot deployment. It eliminates the need to write custom boilerplate for every new robot or policy backend, allowing researchers to deploy, evaluate, and collect data across diverse hardware embodiments using a single framework.

How it works

The system acts as a middleware layer that decouples the robot hardware from the AI policy. It uses a modular architecture consisting of:

• Transports: Supports various communication protocols like ROS 1, ROS 2, and ZeroMQ to talk to different robot hardware.
• Policy Backends: Integrates with multiple inference engines (such as OpenPI, StarVLA, and GR00T) via WebSockets or ZeroMQ.
• Inference Strategies: Implements different execution methods (sync, async, RTC) to handle latency and ensure smooth robot motion.
• Web Console: A browser-based interface for manual teleoperation, data collection, real-time debugging, and evaluation result visualization.

Who it’s for

It is designed for robotics researchers and engineers working on embodied AI, specifically those developing Vision-Language-Action (VLA) models or Reinforcement Learning (RL) policies that need to be tested and iterated on real hardware.

Highlights

• Broad Hardware Support: Compatible with a wide range of robots, including AgileX, ARX, Franka Panda, and AgiBot humanoids.
• Integrated Data Pipeline: Direct teleoperation capture into the LeRobot dataset format with built-in quality control (QC) flags.
• Comprehensive Eval Suite: Supports multi-checkpoint sweeps, prompt shuffling, and synchronous replay of rollout data (video, 3D scenes, and state charts).
• Human-in-the-Loop (HIL): Allows human intervention during policy rollouts to capture correction data for RL fine-tuning.
• Agent-to-Hardware Interface: Includes a preview of EVA-MHS, an MCP-based service that lets coding agents inspect robot state and call policy models.