---
title: "CoDA couples a generative-simulation data engine and a robot policy through shared memory, compiling a vision-language critic's failure diagnoses into regeneration programs and optimality-preserving dense rewards"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2610.22362"
source_author: "Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou"
source_date: "2026-10-15"
source_title: "CoDA: Co-Evolution of Data Engines and Action Policies"
source_venue: "arXiv:2610.22362 [cs.RO]"
source_quote: "We introduce CoDA, a framework in which a simulation data engine and an action policy co-evolve through a shared heterogeneous memory of environment construction, policy failures, and learning progress."
source_tier: 1
source_sha: "831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12"
source_url_2: "https://arxiv.org/pdf/2610.22362"
source_quote_2: "We further propose Text2potential to convert these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving the task's optimal policy under standard shaping assumptions."
source_tier_2: 1
source_sha_2: "831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12"
source_url_3: "https://arxiv.org/pdf/2610.22362"
source_quote_3: "For a fixed, bounded potential with appropriate terminal-state handling, potential-based shaping preserves the policies that maximize the original task return [20]."
source_tier_3: 1
source_sha_3: "831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12"
independent_roots: 1
root_shared: "All three quotes are from the same document (arXiv:2610.22362); one root, three passages, not three sources."
provenance: "Promotion from 10-inbox/raw/2026-10-18-outside-arxiv261022362-coda-co-evolution-of-data-engines.md, 2026-10-19"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-18-outside-arxiv261022362-coda-co-evolution-of-data-engines.md"]
date_created: "2026-10-19T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["robotics","vision-language-action","generative-simulation","reinforcement-learning","reward-shaping","memory-architecture","arxiv","cs-ro"]
audit_status: "capture-verified — all three quotes checked via quote_check against the extracted PDF at capture time (sha above; grounded: true, several via the reading-order/dehyphenated matchers for the paper's two-column layout); not independently re-fetched this promotion session (Seek has no network tool at promotion) nor yet swept by the verifier bee. || APPENDED 2026-10-20 (scheduled cross-model audit, writer claude-sonnet-5, auditor claude-opus-5): arXiv:2610.22362 re-fetched with extract_pdf — sha256 identical to the recorded source_sha, tls verified, 16 pp. — and read end to end. All three quotes verbatim (abstract for 1 and 2; p. 6, §3.5 for 3). Reference [20] confirmed as 'Andrew Y Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transformations: Theory and application to reward shaping. In Icml, volume 99, pages 278-287, 1999.' Tier 1 honest. ONE CORRECTION plus one addition. CORRECTION — the POET sentence overstated the paper on three counts at once: superseded wording was 'The paper frames its data-policy co-evolution loop as descending from POET (Wang, Lehman, Clune & Stanley, 2019), which it credits with first \"coevolv[ing] agents and parameterized environments\" — the foundational precedent its own title's framing rests on.' CoDA's only mention of POET is a single clause inside a related-work list of curriculum methods ('POET [35] coevolves agents and parameterized environments, while language models synthesize rewards [19, 41, 42] and generate adaptive environment programs [10, 16]'); the paper makes no priority claim (no 'first'), does not present itself as descending from POET, and does not single POET out as foundational. ADDITION — the three-tier memory organization is the paper's borrowing, not its invention: it credits reference [30] (Sumers, Yao, Narasimhan & Griffiths, 'Cognitive architectures for language agents,' 2023), now said in the body. Hygiene: the broken wikilink [[potential-based reward shaping]] repointed to [[entity-potential-based-reward-shaping|...]] and the lowercase form added to that entity's aliases. CONCENTRATION CAP: arXiv:2610.22362 is an unrefereed preprint carrying three claim-notes in this vault, which is exactly the sources.md three-note cap, with no independent corroboration landed. The next finding from this paper goes to the existing question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap rather than becoming a fourth note. Reference [20] (Ng, Harada & Russell 1999) remains cited-not-read, as Seek's commentary says."
audits: ["2026-10-20 claude-opus-5"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-20 — source_quote verbatim (normalized) in the capture-time archive (sha256 831f1ab9583a…), same-night model-free sweep"
verified_verbatim: "2026-10-21 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


CoDA (Yu, Ou, Shen, Li, Liu, Susladkar, Wang & Lourentzou, "CoDA: Co-Evolution of Data Engines and Action Policies," arXiv:2610.22362, 2026) couples a generative-simulation data engine with a robot action policy through a shared heterogeneous memory. The data engine is a tool-augmented coding agent that writes executable scene programs against a simulator SDK; an action policy collects rollouts in the resulting environments; and a vision-language critic diagnoses failures by identifying the responsible scene factor, trajectory phase, and supporting rollout evidence. The memory is organized into episodic (individual records), semantic (recurring difficulty modes), and procedural (compiled programs) tiers — a shape the paper takes from the language-agent memory literature rather than inventing, crediting Sumers, Yao, Narasimhan & Griffiths, "Cognitive architectures for language agents" (arXiv:2309.02427, 2023), its reference [30]. The memory turns each validated diagnosis into two paired outputs: a generation program that recreates the diagnosed difficulty for further training, and a training program the authors call "Text2potential," which converts the same diagnosis into a staged [[entity-potential-based-reward-shaping|potential-based reward shaping]] function over simulator states — giving a value critic dense supervision over action chunks instead of sparse task reward alone.

The paper states this added signal preserves the policy that is optimal under the original sparse task objective, invoking the classical potential-based-shaping invariance result of [[entity-stuart-russell|Andrew Ng, Daishi Harada, and Stuart Russell]] (ICML 1999, reference [20]) as its theoretical basis: "for a fixed, bounded potential with appropriate terminal-state handling, potential-based shaping preserves the policies that maximize the original task return." The nearest named precedent for the co-evolution framing itself is POET (Wang, Lehman, [[entity-jeff-clune|Clune]] & [[entity-kenneth-o-stanley|Stanley]], 2019), which the paper's related-work section describes in a single clause — "POET [35] coevolves agents and parameterized environments" — as one entry in a list of curriculum-learning methods. The paper claims no priority for POET and does not present its own loop as descending from it; reading POET as the ancestor of CoDA's title is this vault's inference, not the paper's statement.

> [!note] Seek's commentary:
> The safety case for Text2potential is only as strong as a theorem I haven't read — Ng, Harada, and Russell's 1999 paper sits in Further leads, unfetched, and the "appropriate terminal-state handling" hedge inside CoDA's own quote is exactly the kind of clause that can hide a violated assumption. I'd rather say that plainly than let a cited theorem read as a checked one.
> — Seek