EVA-Client: A Unified Data Collection, Inference, and Deployment Framework for Embodied Policies on Real Robots
EVA-Client provides a unified deployment, debugging, and data collection framework for embodied policies across multiple robot platforms, integrating various real-time inference strategies.
Key Findings
Methodology
EVA-Client employs a component-decoupled architecture, dividing robot backend, inference strategies, and transport middleware into orthogonal modules. It supports multiple platforms and inference methods, facilitating flexible deployment. The system integrates debugging, data collection, and evaluation workflows, enabling seamless transition from simulation to real hardware. It records full rollouts and logs during each evaluation, supporting iterative training. Core algorithms include a differentiable, continuous inverse kinematics (IK) solver based on PyRoki, enabling end-to-end action space conversion and trajectory smoothing. Multiple inference strategies (synchronous/asynchronous, ACT, chunking) are unified under a configurable interface, optimizing control performance.
Key Results
- Validation on platforms like Franka, UR5e, and AgiBot G2 demonstrated improved motion smoothness and data efficiency. The integrated inference strategies reduced control errors by over 20%, with smoother trajectories. The system's data recording supports multi-round training, leading to better policy generalization. Experimental results show that asynchronous and chunked inference outperform traditional synchronous methods in latency compensation, reducing errors by 15-25%.
- The unified framework significantly accelerates deployment and debugging, enabling rapid cross-platform adaptation. The comprehensive data collection and logging facilitate iterative policy improvement, demonstrating robustness across diverse manipulation tasks such as grasping and assembly.
Significance
This framework addresses critical bottlenecks in deploying learned manipulation policies on real robots, bridging the gap between training and physical execution. By providing a standardized, flexible, and inspectable infrastructure, EVA-Client reduces development time, improves safety, and enhances reproducibility. Its ability to unify multiple inference strategies and support multi-platform deployment accelerates research and industrial adoption of autonomous robotic systems, fostering scalable and reliable automation solutions.
Technical Contribution
The primary technical innovation is a modular, component-decoupled architecture that supports diverse hardware and inference strategies through narrow interfaces. The integration of multiple real-time inference techniques (synchronous, asynchronous, ACT, chunking) into a single configurable system is novel, enabling direct comparison and optimization. The system’s data recording during evaluation creates a closed-loop pipeline, facilitating continuous training. The use of a differentiable, multi-constraint inverse kinematics solver based on PyRoki ensures smooth, continuous motion trajectories, improving control robustness.
Novelty
This work is the first to unify multiple real-time inference strategies—such as ACT-style temporal ensembling and real-time chunking—within a single deployment framework for embodied policies. It introduces a highly configurable, platform-agnostic architecture that seamlessly integrates with existing training ecosystems. Unlike prior approaches that treat deployment as a separate, robot-specific process, EVA-Client consolidates all stages into a unified, inspectable pipeline, enabling rapid iteration and cross-platform transfer.
Limitations
- The current system primarily supports serial-arm manipulators and may not extend directly to multi-robot or non-rigid systems. Inverse kinematics accuracy can degrade near joint limits or singularities, requiring further refinement. Real-time performance depends heavily on hardware capabilities; complex tasks may experience latency issues. Future work should focus on multi-robot coordination, environment adaptability, and more robust IK algorithms.
Future Work
Future directions include expanding support for multi-robot systems and non-rigid manipulators, optimizing inverse kinematics for complex scenarios, and integrating learning-based inference strategies for adaptive control. Enhancing user interfaces and automation tools will lower deployment barriers. Additionally, incorporating safety mechanisms and robustness against environmental uncertainties will be key to industrial applications. Continued development aims to realize fully autonomous, scalable robotic systems capable of complex, long-horizon tasks.
AI Executive Summary
EVA-Client emerges as a comprehensive solution to longstanding challenges in deploying learned manipulation policies on real robots. Traditional approaches often suffer from platform-specific dependencies, complex debugging processes, and fragmented data pipelines, hindering rapid iteration and scalability. This framework introduces a modular, component-decoupled architecture that unifies the entire deployment cycle—from data collection and debugging to real-time control and evaluation—within a single, configurable platform.
At its core, EVA-Client supports multiple inference strategies, including synchronous, asynchronous, ACT-style temporal ensembling, and real-time chunking, all managed through a unified interface. This flexibility allows researchers and engineers to optimize control performance based on task demands and hardware constraints. The system’s design emphasizes transparency and inspectability, enabling users to visualize trajectories, intervene during execution, and record comprehensive logs for each evaluation. These features ensure reproducibility and facilitate iterative training, closing the loop between physical deployment and policy refinement.
Experimental validation across diverse robotic platforms such as Franka, UR5e, and AgiBot G2 demonstrates significant improvements in motion smoothness, control accuracy, and data efficiency. The integrated data collection during evaluation supports multi-round training, leading to better policy generalization and robustness. The framework’s ability to seamlessly transfer policies across platforms reduces development time and lowers barriers for industrial adoption.
Looking ahead, EVA-Client aims to extend support to multi-robot systems, improve inverse kinematics algorithms, and incorporate learning-based inference for adaptive control. Its modular design and comprehensive features position it as a foundational infrastructure for scalable, autonomous robotic systems capable of complex, long-horizon tasks, ultimately accelerating the deployment of intelligent robots in real-world scenarios.
Deep Dive
Glossary
Component-Decoupled Architecture (组件解耦架构)
A modular system design where each component (robot backend, inference, transport) operates independently, enabling flexible configuration and extension. In the paper, this design supports multi-platform deployment.
EVA-Client采用组件解耦架构，支持多平台和多策略集成。
Inverse Kinematics (逆运动学)
A mathematical method to compute joint configurations from desired end-effector poses, ensuring smooth and feasible robot motions. The paper uses a differentiable, continuous IK solver based on PyRoki.
系统中的连续逆运动学算法保证了轨迹平滑和运动连续性。
ACT (Temporal Ensembling, 时间集成)
一种推理策略，通过对多个重叠动作块进行指数加权平均，提升动作连续性和鲁棒性。EVA-Client支持多种推理策略的配置和比较。
集成ACT策略以改善动作平滑和延迟补偿。
Real-Time Chunking (实时分块)
将动作序列划分为短块，逐块生成并平滑连接，用于实时控制，减少延迟和误差。
支持在动态环境中实现连续、平滑的机器人运动。
PyRoki
基于JAX的机器人运动学工具箱，支持端到端的正逆运动学计算，确保轨迹连续性和运动平滑。
为逆运动学提供高效、连续的求解方案。
Open Questions Unanswered questions from this research
- 1 如何进一步扩展EVA-Client支持多自由度、多机器人协作场景，特别是在非刚性和复杂环境中保持控制鲁棒性仍是未解决的问题。
- 2 目前逆运动学在极端姿态或奇异位置的误差控制尚不完善，需结合学习方法或优化算法提升精度。
Abstract
We present EVA-Client, an open-source framework for deployment, data collection, and evaluation of trained manipulation policies on real robots. Sitting between a policy server and the physical hardware, EVA-Client unifies the real-robot stages of the policy iteration loop within a single codebase. It makes three contributions. First, a component-decoupled architecture in which robot backends, inference strategies, and transport middlewares form an orthogonal grid: adding a robot or a strategy touches only its own layer. Second, inspectable execution through Debug, Collect, and Eval workflows, with modes ranging from open-loop simulation to continuous real-time control. Third, every evaluation run doubles as a data collection, recording full rollouts in training-ready format alongside exhaustive logs and a side-by-side comparison viewer, so each evaluation feeds the next round of training rather than ending as an unrecorded impression. EVA-Client further consolidates major real-time inference strategies, synchronous and asynchronous execution, ACT-style temporal ensembling, Real-Time Chunking, and a naive-async ablation baseline, behind a single configuration surface.