EVA-Client: A Unified Framework for Deployment, Evaluation, and Data Collection on Real Robots One policy, any robot — the smooth all-in-one real-robot stack. Debug, record, evaluate, visualize, all in the browser. EVA-Client driving an AgileX bimanual arm end-to-end from the browser — teleop → record → π₀ checkpoint → smooth async deploy. Real hardware, not a rendering. Jump to: What's NEW · What you get ·…
EVA-Client: A Unified Framework for Deployment, Evaluation, and Data Collection on Real Robots
One policy, any robot — the smooth all-in-one real-robot stack. Debug, record, evaluate, visualize, all in the browser.
EVA-Client driving an AgileX bimanual arm end-to-end from the browser — teleop → record → π₀ checkpoint → smooth async deploy. Real hardware, not a rendering.
[2026-08-28] EVA-Client 0.2.0 EVA-Client now supports WebXR-based VR controller teleoperation, ARX X5 hardware, dual-YAM leader-follower data collection, export to LeRobot v3.0, MCAP, and HDF5, and multiple coding-agent skills for adding robots, calibrating cameras, and debugging robot SDKs.
[2026-08-28] EVA-MHS Preview. Inspired by Anthropic’s Model Hardware Standard (MHS) research preview, EVA-MHS explores an MHS-style agent-to-hardware interface for robotics. Its MCP service allows coding agents to inspect runtime state, call policy models, and directly control robots. EVA-MHS is available on the dev/mcp_preview branch and will be formally released in a future update.
[2026-07-22] New RL Workspace. EVA-Client now supports human-in-the-loop (HIL) intervention during policy rollouts and real-time value-curve visualization. See the RL workspace and HIL intervention demo.
[2026-07-18] Headless CLI + ZMQ control channel. Drive the full deploy / eval / RL / collect loop from a single .py config without the browser — every console button (incl. HIL takeover) mirrored on a ZMQ socket. See docs/console.md and docs/control-channel.md.
🚀 Deployment. One command brings up a real-robot closed loop:
.py config → transport (ROS1 / ROS2 / ZeroMQ / offline dataset) → policy
backend (OpenPI, OpenPI-RTC, StarVLA, GR00T, mock, replay) → inference
strategy (sync / async / naive / ACT-ensemble / RTC) with live latency
compensation. 8 robots already — joint-space or EEF-space (PyRoki IK),
all live-switchable from the DEBUG tab.
📊 Evaluation. Multi-checkpoint sweeps with per-trial records: every
rollout captures camera video, 3D URDF scene, per-dimension state charts,
and per-prompt milestone scores into dataset metadata, then replays
synchronously in the RESULT tab. Prompt shuffling, and remote
policy servers over SSH port-forward are first-class.
🎥 Data collection. Teleop capture straight into LeRobot v2.1 episodes
from the COLLECT tab — background saver, in-tab QC PASS/FAIL replay, camera
streams encoded to mp4, per-frame green/red quality flags. Teleop demos and
model rollouts share one on-disk layout.
Tests, lint, type-check, and how the robot side is faked
🗺️ Roadmap
More robots. Extend the robot zoo to more embodiments — dual-arm
manipulators (Tianji, …), humanoids
(Unitree H1/G1, Fourier GR-1, Booster T1, …) and mobile / wheeled
platforms (mobile ALOHA, Galaxea R1 base, quadruped + arm).
Human-in-the-loop data collection for RL. Interventions during
policy rollout captured as preference / correction data, DAgger-style
relabeling, and reward-model signals piped back through the LeRobot
episode format for online RL fine-tuning.
Embodied agent. Wrap the deployment + evaluation loop with a
language-driven planner (VLM / VLA + tool use) so long-horizon tasks
can be decomposed, executed, verified, and re-planned end-to-end from
the same console.
Data annotation. Extend Collect mode with fine-grained task and
sub-task annotation, segmenting long-horizon episodes into labeled
sub-task units and milestones within the same LeRobot dataset. This
makes a collection reusable at the level of individual manipulation
phases.
📝 Citation
If EVA-Client is useful for your research or product, please cite:
@misc{yang2026evaclient,
title={EVA-Client: A Unified Data Collection, Inference, and Deployment Framework for Embodied Policies on Real Robots},
author={Heqing Yang and Yang Yi and Liyao Wang and Linqing Zhong and Donglin Yang and Ruipu Wu and Zitong Bai and Fengjiao Chen and Manyuan Zhang and Linjiang Huang and Si Liu},
year={2026},
eprint={2607.02646},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2607.02646},
}
📬 Contact
For questions or collaboration, feel free to reach out via WeChat:
License & Acknowledgements
This project is licensed under the Apache-2.0 License. See LICENSE
for more information.
Builds upon several excellent open-source efforts, including
PyRoki and
jaxls for kinematics,
OpenPI for policy serving,
the LeRobot dataset format, and a
vendored mmengine-style config system from
MMEngine.
评论