零次方 ZERITH 研究 ESI

Technical report/November 15, 2026

ESI: An Embodied Self‑Improvement Flywheel for LLM–VLA Agents

ESI turns a pretrained LLM and a pretrained VLA into one coordinated robot agent, by learning from their own interaction.

How ESI worksSchematic
How ESI works, in three acts Before ESI, the VLA goes its own way instead of following the trajectory the LLM intends, and stops short of the target. The flywheel then updates the prompt, the skill interface and the VLA weights. After ESI, the VLA follows the same intended trajectory to the target. LLM plans and reviews VLA acts ? LLM intent VLA trajectory Prompt Skill VLA weights

Agent + VLA

Ahead of all published LLM+Model embodied agents, without generating task-specific control code.1

Ten-task mean

The problem

Why connecting an LLM and a VLA is not enough

A reasoning model and an action model can each be capable and still fail together. The failures sit at the interface between them.

Intent gap

The LLM chooses a goal. The VLA does something familiar instead.

LLMLeft arm: Grasp red block.
VLAApproaches another object, or closes before reaching a good contact.

Handoff gap

The LLM repairs the motion and gives control back. The VLA cannot continue from there.

LLMCorrects the approach and restores a promising grasp pose.
VLARepeats its original motion from a state it has never trained on.

More LLM control fixes the episode. It does not fix the next one. ESI treats these interactions as training data.

Results

Higher task success with less LLM control

The evaluation uses ten of the most challenging tasks in RoboDojo.1 Against a matched baseline on the same layouts, ESI completes more tasks while the LLM takes direct control far less often. It also exceeds every published reference in the table below on both metrics.3

Against published references

All results are in simulation: RoboDojo, with a dual-arm ARX X5 robot. The first three columns are ESI and its two components evaluated alone. The other columns are entries on the RoboDojo leaderboard for the same task names. Type is the entry's own category on the leaderboard; GO-3 Nano Preview, for example, is listed as Agent + WAM.1

Ten tasks in one model

The most challenging RoboDojo tasks, all run by one LLM and one VLA.4

VLALLMPath so farNext 25 steps

Same outcome, less help

LLM control where the baseline leaned on it most

VLALLM
BaselineLLM executes 19.26% of all steps
ESI11.12%

Paired cases ranked by how much of the episode the baseline ran under direct LLM control; only the cases where that share is highest are shown. Each line is one episode, scaled to its own length.

How it works

One loop that improves three components

Inside an episode, the LLM and the VLA plan, propose, review and act. Across rounds, the same record updates the prompts that guide the LLM, the skill interface that carries its intent, and the VLA that executes it. The LLM's weights stay fixed.5

How a guidance update moves through the system

Illustrative example

After a round, a proposed rule must cite the failures behind it, a recovery that was seen to work, and counterexamples that limit it.

Recorded cases

Proposed rule

The path of an accepted rule

The keyboard lies rotated beside its frame, a gripper at its edgeMany pushes later the keyboard has shifted and is still rotated
Pushed repeatedly, still rotatedOne of two cited failures
When it applies

When repeated supported keyboard pushes fail to improve alignment, distinguish missed edge engagement from engagement producing the wrong motion.

The keyboard is still at an angle under the armThe keyboard now lies nearly square on the desk
A new contact turns it squareOne of two cited recoveries
What to do

If the fingers cross the border while the keyboard stays still, clear and adjust the approach height to engage an accessible edge. If the keyboard moves but mostly translates while the required rotation stalls, change the contact location rather than extending the same push. Use the resulting keyboard motion to choose further corrections.

Both grippers at the edges of a rotated keyboardThe keyboard sits square, both grippers clear of it
The VLA's own pushing worksOne of three cited counterexamples
What to leave alone

Preserve productive [VLA] pushes, including differential pushing; neither a rejected approach nor remaining offset alone requires changing contact.

Automatic checks
  • Two or more independent cases
  • A working recovery and counterexamples
  • Within 180 words
  • An identifier of its own
Independent review

A separate conversation reads the cited decisions and images itself. It approves, rejects, or asks for a closed-loop test.

Next round's prompt
organize_tableTask goal and success conditions, fixed
learned_keyboard_contact_responseWhen repeated supported keyboard pushes fail to improve alignment, …

Published as one hash-pinned bundle. The next collection round loads exactly that file.

The frames and the rule come from the system's analysis of round-one Organize table episodes; the bracketed word replaces a term from the record. The right-hand column illustrates the path of a rule that passes. The reported results use fixed guidance.5

Shared interface

The LLM's instruction is also the VLA's training label

The subtask the LLM sends online is the same language the VLA is trained on offline. Build one and see exactly what the VLA receives.

Build a subtask

In an episode. The VLA receives this text next to its camera images and joint state.
In training. The same sentence is the language label of a kept window.

The catalog. One shared vocabulary for one VLA; each entry describes an operation, and none is a separate controller. Pick one to build it above.

Data pipeline

How interaction records become training data

Follow one failed episode from the simulator into the training set. Each step is a design choice, not a separate ablation.

Component study

What each component contributes

Roll back one adapted component at a time, on 100 matched cases.7

By task

Ten cases per task and configuration. Darker cells are higher.

Unseen task

A task outside the training data

Complete the equation was never used in flywheel collection or post-training. ESI solves four of the five test layouts, the baseline two, and LLM-executed steps fall from 99.0 to 43.6 per episode.8

Intent following

One scene, three requested orders

One scene, one target arrangement. Only the requested pad order changes.8

ESI completes all three orders. The baseline completes two.

Why it matters

A path from better reasoning to better robot execution

The LLM works twice. It helps finish the task now, and its decisions and corrections become supervision for the policy. A stronger reasoning model is therefore a second way to improve the robot, next to more robot experience. This is a direction the framework opens, not a measured result.

  1. The ten tasks are chosen to be among the most challenging in RoboDojo: the suite starts from the GPT-as-Policy ten-task suite and replaces several easier tasks with more demanding ones. Twenty scheduled cases each. Score and success rate are task means with tasks weighted equally. Reference values come from the RoboDojo leaderboard for the same task names, using each entry's reported case set. The strongest reference Score is 48.51 (WorldScape-R²SI) and the strongest reference success rate is 38.60% (HexaShell-Preview); ESI exceeds them by 13.52 Score and 10.81 points. Code-based methods generate task-specific control code; ESI does not.
  2. Environment steps executed under direct LLM control, averaged per episode. An LLM decision executes at least one such step.
  3. The matched baseline is GPT-as-Policy with the original RoboDojo-adapted VLA, task-only policy conditioning and baseline decision guidance. Both agents use GPT-6-Astra with the same environment code, layouts, scene seeds, motion executor and native step budgets. Three native-invalid layouts are excluded from both, leaving 197 paired cases. The comparison measures the two deployed systems as a whole.
  4. For each task, the ESI episode is the ESI success whose LLM-executed steps are the median among that task's ESI successes (the lower middle one when even); Fill pen holder has one, Sweep blocks and Organize table two each. The baseline panel shows the matched baseline episode on the same layout and split. All ten ESI episodes share one LLM and one VLA checkpoint. Row numbers cover all of the task's paired cases.
  5. The three-round configuration behind the reported numbers updates the VLA each round with a fixed prompt and skill catalog. The automatic prompt and skill update mechanisms are integrated in the released system.
  6. Of the candidate training-window starts in each round's 200 collected episodes, 47.9%, 48.6% and 49.7% come from episodes that ended in failure. This counts window starts, not independent trajectories or sampling probability.
  7. Ten tasks with ten matched cases each. The ESI reference is the matched subset of the main evaluation; the rollback runs use a later evaluation implementation, so each configuration is reported as an absolute result. Three-hour timeouts count as non-success.
  8. In the transfer and order studies, both agents use subtask feedforward and action review; the baseline has the initial policy, free-form subtasks and no added task guidance. The equation task was excluded from flywheel collection and post-training. The five layouts were chosen for follow-up after preliminary testing on five others. Each order was run once per agent in one fixed scene, and order adherence was checked by video inspection.

Technical report, November 15, 2026

ESI: An Embodied Self-Improvement Flywheel for LLM–VLA Agents

Yundi Cai, Zihan Wang, Chenxi Han, Yi Cheng, Houde Liu, Zilei Wang

University of Science and Technology of China, Tsinghua Shenzhen International Graduate School, ZERITH Robotics

@article{cai2026esi, title = {ESI: An Embodied Self-Improvement Flywheel for LLM--VLA Agents}, author = {Cai, Yundi and Wang, Zihan and Han, Chenxi and Cheng, Yi and Liu, Houde and Wang, Zilei}, journal = {arXiv preprint arXiv:2611.18176}, year = {2026}, eprint = {2611.18176}, archivePrefix = {arXiv}, primaryClass = {cs.RO}, url = {https://arxiv.org/abs/2611.18176} }