EMBODIED TASK AGENT

OpenETA

An Embodied Task Agent Framework for
Trustworthy Physical Closed Loops

OpenETA Team · Shanghai Innovation Institute

Code Technical Report Robot Stack OpenETA working trace
OpenETA closed-loop framework overview

Why OpenETA

Task intelligence needs a trustworthy loop.

Action prediction alone does not tell an embodied agent whether a grasp succeeded, whether an object remains attached, or whether a timed-out request changed the world. OpenETA transfers the digital-agent loop—understand, invoke a Tool, inspect feedback, and decide again—into simulators and robots.

OpenETA separates cognition from execution authority across three roles: the Agent proposes structured intent, the Interface validates contracts and gates side effects, and the World returns trusted state evidence. World-changing actions create a fresh-observation obligation before another state-dependent decision.

Approach

Closed-loop Tool use, not a one-shot control program.

01
Observe

Normalize multimodal state and provenance.

02
Decide

Select one bounded Tool from current evidence.

03
Act

Validate authority, safety gates, and side effects.

04
Verify

Read a trusted receipt and observe the changed world.

Comparison of VLA, WAM, code-policy, and OpenETA control loops
OpenETA re-enters the Agent after every bounded Tool call, increasing feedback frequency and localizing failures to a specific stage.

Evidence Chain

A successful API call is not task success.

OpenETA evidence chain from target selection to task checker
OpenETA separates call acceptance, motion completion, physical state change, and task-level verification. Receipts and observations remain replayable.

Simulation Evaluation

A complete 40-task × 10-seed LIBERO baseline.

The frozen descriptive baseline runs without task-specific policy training. Official LIBERO reward is the only success signal; human and agent assistance are zero for all 56 successful episodes.

SuiteSuccessesRate
LIBERO Spatial8 / 1008%
LIBERO Object26 / 10026%
LIBERO Goal21 / 10021%
LIBERO-101 / 1001%
Overall56 / 40014.0%

18/40 tasks succeed at least once. Three LIBERO-10 cells were affected by simulator-TTL expiration and remain counted as zero, so this is descriptive rather than infrastructure-clean confirmatory evidence.

Dominant terminal outcomes

Episode timeout
215
Unattended ask_human
66
Max turns
35
Status without reward
24
Simulator TTL
3
Remote termination
1

Real Robot

Qualitative UR5e demonstrations.

These retained recordings demonstrate hardware execution stages, not a trial denominator or success rate. Formal primitive- and task-validation remain pending a frozen, trial-level evidence batch.

Four stages of a UR5e sponge-to-tray manipulation demonstration
One retained sponge-to-tray recording: contact approach, lift-probe verification, multi-waypoint transport, and final placement.
Complete qualitative example

Sponge block → tray

The retained view visibly supports contact, attachment verification, transport, and placement.

d23664831877…961785c
Bounded visual claim

Bell pepper → basket

Grasp, lift, and transport are visible. The archived view does not support a final in-basket relation, so this is not presented as complete task success.

90f53fa34971…e19064
Evidence boundary. The webpage videos are lower-resolution, silent previews derived from the retained MP4 files. Their hashes identify the originals; selected successful recordings cannot estimate reliability or replace complete success-and-failure trial coverage.

Self-Evolution

Experience is proposed locally and promoted conservatively.

OpenETA turns traces into candidate strategies, then requires paired replay, non-regression checks, and held-out validation before a candidate can affect future execution.

In the current exploratory study, no candidate passed the promotion gate. This supports the safety mechanism—not a performance-improvement claim—and motivates more identifiable, stage-local operators in the second phase.

OpenETA research roadmap
Only verified experience is eligible to re-enter the deployment loop.

Resources

Read, use, and cite OpenETA.

GitHub Repository Download Report 
@misc{chen2026etanewagenticparadigm,
      title={ETA: A New Agentic Paradigm for Embodied Tasks}, 
      author={Yitong Chen and Zezheng Huai and Sixian Li and Yubang Wang and Haozhe Zhang and Yifei Zhang and Hechang Chen and Jingjing Gong and Yu-Gang Jiang and Xipeng Qiu},
      year={2026},
      eprint={2608.03924},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2608.03924}, 
}

The author list is provisional and will be updated with the frozen release metadata.