Executable coordination for embodied teams with distributed private information.
arXiv · PDF · LaTeX Source · Evidence · Reproducibility · Citation
Peng He1,*, Junning Zhu2,*, Haohan Yuan3, Jianpeng Liang4
* Equal contribution
|
|
|
|
| 1 Tsinghua University |
2 Beijing Normal-Hong Kong Baptist University |
3 University of North Carolina at Charlotte |
4 University of California San Diego |
GenCoord in one view. Distributed private facts become a shared executable commitment, then flow through grounded execution and terminal-state verification.
GenCoord turns distributed private facts into executable, multi-step commitments that preserve their task meaning across communication, resolution, grounding, execution, and verification.
private fact → β(z) → SELF + REQ → resolved commitment → grounded skills → verified joint action
Embodied teams frequently distribute the decisive facts of a joint route across agents: one agent sees the goal, another knows the relevant workcell capability, and the team must bind actor, handoff item, destination, and continuation into one coordinated plan.
GenCoord makes that execution-relevant binding explicit. A local Qwen3.5-0.8B model composes role-local paths in a typed, multi-step SELF+REQ schema. Bounded ACCEPT / REJECT / COUNTER feedback resolves peer-local capability information, and the resolved commitment is parsed, schema-checked, canonically materialized, compiled to Mineflayer skills, executed, and verified through handoff and terminal state.
| Distributed private facts | Executable commitment | Verified joint action |
|---|---|---|
| Goal, capability, inventory, and route evidence stay with the agents that observe them. | One typed SELF+REQ object binds actor, handoff item, destination, and continuation. |
The resolved object drives canonical materialization, skill compilation, grounded execution, handoff, and terminal verification. |
| Capability resolution | Commitment horizon | Communication efficiency |
|---|---|---|
| 50% → 100% | 91.3% → 98.1% | 92.8% less peer traffic |
| Bounded capability feedback closes paired local ambiguity across three independently trained seeds. | Multi-step commitments raise held-out-template success while reducing model decisions from 2.91 → 1.98 per episode (−32%). | Short DSL preserves 128/128 closed-loop semantic clusters and reduces median time-to-commitment by 68.2% versus controlled free-form. |
Paired counterfactual interventions isolate a bidirectional content-to-route mechanism: changing the injected task consequence while holding the world, call schedule, and executor fixed redirects requester revision and receiver execution exactly along the corresponding route.
SELF resource.obtain(q=4,item=oak_planks)
> resource.deliver(q=4,item=oak_planks,to=agent_b)
REQ agent_b craft.item(q=1,input=oak_planks,item=crafting_table)
> resource.deliver(q=1,item=crafting_table,dst=order_chest)
The same schema carries proposal-time routes, resolved role-local obligations, and the executable object consumed by the grounded stack. Short DSL presents this structure directly as the agent-to-agent executable interface.
The repository packages the complete paper, supplementary material, final figures, canonical evidence, and deterministic builders into one research release.
| Asset | Included material |
|---|---|
| Preprint | Combined paper and supplementary PDF. |
| Paper source | Main paper, supplement, references, venue-support files, and generated LaTeX tables. |
| Figures | Final paper figures used throughout the manuscript and supplement. |
| Canonical evidence | Source values, episode metrics, paired comparisons, trace registry, source ledger, and experiment summaries. |
| Deterministic builder | Standard-library Python pipeline for rebuilding and validating the released E1/E2 tables. |
git clone https://github.com/JulianZJN/GenCoord.git
cd GenCoord
# Build the paper and supplement
make main
make supplement
make combined
# Rebuild and verify the canonical evidence
make evidence
make verify-evidenceBuild requirements and outputs
- A recent TeX Live distribution with PDFLaTeX and BibTeX.
- Python 3.9 or newer; the evidence builder uses the Python standard library.
pdfunitefor the combined preprint target.
paper/main.pdf
paper/supplement.pdf
build/GenCoord_preprint.pdf
The build produces a 9-page main paper, a 16-page supplement, and a 25-page combined preprint.
The arXiv-oriented source order is recorded in paper/00README.json.
The deterministic evidence pipeline validates the experimental schema, conditions, episode counts, semantic-cluster unit, training seeds, parse and grounding validity, fallback count, and selected paired-comparison deltas before emitting the released tables and canonical data.
make evidence
make verify-evidenceThe consistency target rebuilds every generated output and verifies exact agreement with the tracked research artifacts. Full commands and generated-file maps are documented in docs/REPRODUCIBILITY.md.
.
├── README.md
├── CITATION.cff
├── Makefile
├── assets/
│ ├── teaser.png
│ └── affiliations/
├── docs/
│ ├── GenCoord_preprint.pdf
│ └── REPRODUCIBILITY.md
└── paper/
├── main.tex
├── supplement.tex
├── references.bib
├── figures/
├── tables/
└── anc/
├── scripts/build_e1e2_tables.py
└── source_data/
The public preprint is available as arXiv:2608.22055. Citation-ready metadata is also provided in CITATION.cff:
@misc{he2026gencoord,
title = {GenCoord: Skill-Path Commitments under Private Information},
author = {Peng He and Junning Zhu and Haohan Yuan and Jianpeng Liang},
year = {2026},
eprint = {2608.22055},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2608.22055}
}GenCoord uses Minecraft as its embodied evaluation environment and compiles validated commitments to Mineflayer skills. Original project code and build glue are MIT-licensed. Manuscript source, figures, evidence, institutional marks, venue-support files, and third-party materials follow the path-level terms recorded in LICENSE, THIRD_PARTY_NOTICES.md, and assets/affiliations/SOURCES.md.