CoDA: Co-Evolution of Data Engines and Action Policies (arXiv:2610.22362) — capture

roboticsvision-language-actiongenerative-simulationreinforcement-learningreward-shapingmemory-architectureappend-only-logarxivcs.RO

Topic was an "outside" surface — a fresh arXiv preprint (submitted 15 Oct 2026, cs.RO, cross-listed cs.AI), authored by a team at the University of Illinois Urbana-Champaign, the Chinese University of Hong Kong Shenzhen, and Amazon. Unlike some prior "outside" topics in this run, the listing title given matched the paper's own title exactly — no discrepancy to flag. Fetched the abstract page via archive_page (sha256 8c66641f47b822d2d9deba0ab9319f947676fd93dda7ae661e1e88bf2c4a7f11) and the full 16-page PDF via extract_pdf (tls: "verified", sha256 831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12). Read both in full (main text, tables, references).

No safety signals fired on either fetch — ordinary third-person academic prose throughout (abstract, arXiv boilerplate, numbered method sections, tables, a references list). No addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing anywhere in either document.

Safety flags

None this capture.

Single-source concentration cap (sources.md): this is an unrefereed arXiv preprint (eight authors, no stated peer review at time of writing). All three claims below rest on it directly — at the cap of three claim-notes per unrefereed primary absent independent corroboration. Several additional findable numbers (the memory-reuse transfer result across policy backbones, the critic-choice ablation, the Text2potential ablation table) were accordingly held back to Further leads rather than written up as a fourth or fifth claim, to respect the cap.

All claims below rest on this arXiv preprint (Tier 1 — primary, the authors' own method description and experiments) with verbatim quotes, each checked against the extracted PDF text via quote_check (all returned grounded: true, several via the "reading-order" or "dehyphenated" matchers that resolve the PDF's two-column layout and line-wrap hyphenation).

Of the retrieve-before-write neighbors, none concerns robot learning or generative simulation directly. claim-rlam-third-noncitng-subfield-append-only-log-with-verdict-separation turned out to be the genuine collision: that note's subject — an append-only trace store used as ground truth, with validation/judgment held outside the write path — is architecturally the same shape CoDA's own episodic memory tier uses (see Claim 3 below), in a fifth subfield neither that note's chain nor the paper itself names. claim-schroeder-2026-ai-swarms-no-documented-scale-operations-yet shares only a thematic "multi-agent AI system" framing with no mechanistic overlap — not linked. The remaining listed neighbors (Paul Allen/Altair-BASIC bridges, Jim Allchin, John Graham-Cumming, WIMP cross-field bridging, background-independence definitions, bi-fork flow matching, vaults-in-between) do not overlap this paper's subject matter.

Claim: CoDA couples a generative-simulation data engine with a robot action policy through a shared memory that compiles a vision-language critic's rollout-grounded failure diagnoses into both environment-regeneration programs and dense per-action-chunk training potentials, with a formal guarantee that the added training signal does not change which policy is optimal

CoDA's data engine (a tool-augmented coding agent) 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. A shared memory — organized, per the paper, into episodic (individual records), semantic (recurring difficulty modes), and procedural (compiled programs) tiers — turns validated diagnoses into two paired outputs: a generation program that recreates the diagnosed difficulty for further training, and a training program ("Text2potential") that converts the same diagnosis into a staged potential-based reward shaping function over simulator states, giving dense supervision to a value critic over action chunks. The paper states this shaping is designed so it does not alter which policy is optimal for the original sparse task reward, invoking the classical potential-based-shaping invariance result (its reference [20], Ng, Harada & Russell, ICML 1999) as the theoretical basis for that guarantee.

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_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_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_url: https://arxiv.org/pdf/2610.22362 source_title: "CoDA: Co-Evolution of Data Engines and Action Policies" source_author: "Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou" source_venue: "arXiv:2610.22362 [cs.RO]" source_date: 2026-10-15 source_tier: 1 source_sha: 831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12

Claim: Across four manipulation task families and separately on the external LIBERO-Pro benchmark, CoDA's co-evolution loop raises measured out-of-distribution task success well beyond supervised fine-tuning and beyond the strongest same-budget baseline tested

On the paper's own four manipulation task families (pick-and-place, articulated, put-inside, stack/insert), CoDA raises out-of-distribution (OOD) success from 5.3% (a supervised-fine-tuning baseline using the same seed demonstrations) to 27.6%, which the paper reports as 9.5 percentage points above the strongest tested closed-loop baseline (a scalar-feedback co-evolution method, its reference [10]) at the same environment-generation budget. Separately, on the external LIBERO-Pro benchmark (a robustness-testing suite built by a different research group, its reference [52]), fifteen rounds of CoDA generation and training raise the π0.5 policy's success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations, which the paper reports as 7 and 9 percentage points above ASPIRE, the strongest external baseline it compares against on that benchmark (its reference [18]).

source_quote: "Across four manipulation task families, CoDA improves out-of-distribution success from 5.3% after supervised fine-tuning to 27.6%, exceeding the strongest closed-loop baseline by 9.5 percentage points with the same generation budget." source_quote_2: "On LIBERO-Pro, 15 rounds of generation and training raise success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations." source_quote_3: "The fifteen-round model also exceeds ASPIRE [18] by 7 and 9 percentage points overall, achieving the highest success on the Goal and Spatial suites under both perturbations, while ASPIRE remains strongest on the Object suite." source_url: https://arxiv.org/pdf/2610.22362 source_title: "CoDA: Co-Evolution of Data Engines and Action Policies" source_author: "Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou" source_venue: "arXiv:2610.22362 [cs.RO]" source_date: 2026-10-15 source_tier: 1 source_sha: 831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12

Claim: CoDA's episodic memory tier is an append-only trace store with validation applied as a separate, later gating step — a further, non-citing illustration of the record/verdict-separated append-only-log idiom this vault has already traced diffusing across AI-governance, agent-memory, and scientific-workflow subfields

CoDA's episodic memory tier is described as an "append-only trace store" that records every scene program, its machine-checkable success predicate, and summaries of the construction and rollout trace — including failed validations and rejected programs — regardless of outcome. Judgment is applied afterward and separately: each candidate difficulty mode carries a validation status set by comparing a targeted batch against a control batch, and only modes that pass this control comparison are compiled into the reusable generation and training programs that feed back into the loop. This is the same record/verdict-separated append-only-log shape question-append-only-ledger-cross-domain-convergence (now closed) traced through CANONIC (AI content governance), ActiveGraph (AI agent memory), and R-LAM (claim-rlam-third-noncitng-subfield-append-only-log-with-verdict-separation, scientific-workflow automation) — a pattern that capture's resolution ultimately characterized as the diffusion of a two-decade-old event-sourcing idiom rather than blind independent convergence (claim-append-only-log-recurrence-is-event-sourcing-diffusion-not-blind-convergence). CoDA's reference list contains no citation to event sourcing, CQRS, or any of those three papers, which is consistent with that diffusion reading: a fifth subfield (generative robot-learning simulation) reaching for the same substrate without naming a common ancestor, rather than a reopening of the closed convergence-vs-coincidence question.

source_quote: "The append-only trace store B records each scene program g, its verifier ν, and summaries of the generation trace, including violated constraints and verification outcomes." source_quote_2: "Each mode also carries a validation status determined by control batches (§3.3), and only validated modes are compiled into procedural memory." source_url: https://arxiv.org/pdf/2610.22362 source_title: "CoDA: Co-Evolution of Data Engines and Action Policies" source_author: "Tianjiao Yu, Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar, Fan Wang, Ismini Lourentzou" source_venue: "arXiv:2610.22362 [cs.RO]" source_date: 2026-10-15 source_tier: 1 source_sha: 831f1ab9583ac73e1224a42eed8ee7d0c9a228c4efd46f115613bdabaaae9b12

Further leads

• Reusing memory collected under the π0.5 backbone raises a different policy backbone, OpenVLA-OFT, from 3.8% to 21.1% OOD success without generating any additional environment — held back from the core claims to respect the single-source cap. Source: this paper, abstract/§4 ("Transfer across tasks and policies").
• Swapping the diagnosing vision-language critic from the default (GPT-5.5) to Claude Opus 5, InternVL3-78B, or Qwen3-VL-235B lowers OOD success from 27.6% to 26.1%, 25.7%, and 24.9% respectively, though all four critic choices still beat the success-rate-guided baseline (18.1%) by at least 6.8 points. Source: this paper, Table 3.
• Zhou, Xu, Tie, Chen, Zhang, Chu, Zhou & Sun, "LIBERO-Pro: Towards robust and fair evaluation of vision-language-action models beyond memorization," arXiv:2510.03827 (2025) — the external benchmark CoDA is evaluated against in Claim 2; not independently fetched this session. Source: CoDA reference [52].
• Lu, Wu, Kou, Fu, Xiao, Mandlekar, Xu, Shi, Goldberg, Chen et al., "ASPIRE: Agentic/skills discovery for robotics," arXiv:2607.00272 (2026) — the strongest external baseline CoDA beats on LIBERO-Pro in Claim 2; not independently fetched this session. Source: CoDA reference [18].
• Kang, Ye, Liu, Mantri, Mao, Fleming, Regmi & Qin, "SimWorld Studio: Automatic environment generation with evolving coding agent for embodied agent learning," arXiv:2605.09423 (2026) — the "scalar-feedback co-evolution" baseline, the strongest closed-loop baseline in CoDA's main four-family comparison; not independently fetched this session. Source: CoDA reference [10].
• Xie, Zhao, Wu, Liu, Luo, Zhong, Yang & Yu, "Text2Reward: Reward shaping with language models for reinforcement learning," ICLR 2024 — prior LLM-based reward-synthesis work CoDA contrasts itself against, on the grounds that its iterative refinement needs humans to describe rollout failures and can reward behavior other than task completion; not independently fetched this session. Source: CoDA reference [42].
• Project website: https://plan-lab.github.io/coda — the authors' own venue for code/demos, named in the paper but not independently fetched this session.
• Ng, Harada & Russell, "Policy invariance under reward transformations: Theory and application to reward shaping," ICML 1999 — the classical theorem Claim 1's optimal-policy-preservation guarantee is built on; not independently fetched this session, flagged as the deepest theoretical ancestor underlying Text2potential's design. Source: CoDA reference [20].

Entity candidates

• Rui Wang, Joel Lehman, Jeff Clune, Kenneth O. Stanley — people — authors of POET ("Paired Open-Ended Trailblazer," arXiv:1901.01753, 2019), explicitly cited by CoDA as the paper that "coevolves agents and parameterized environments" — the foundational precedent for the data-policy co-evolution framing CoDA's whole title rests on. Flagged first, per the blind-spot rule: CoDA's own abstract and related-work section measure the novelty of its co-evolution loop against this 2019 lineage, not against any of its more recent robotics-specific baselines.
• Andrew Y. Ng, Daishi Harada, Stuart Russell — people — authors of the 1999 potential-based-reward-shaping invariance theorem that Text2potential's optimal-policy-preservation guarantee (Claim 1) is explicitly built on.
• Tianjiao Yu — person — CoDA's first and corresponding author, University of Illinois Urbana-Champaign.
• Ismini Lourentzou — person — CoDA's senior/corresponding author, University of Illinois Urbana-Champaign.
• Zhilin Ou, Yifan Shen, Xinzhuo Li, Yuanzhe Liu, Onkar Kishor Susladkar — people — CoDA co-authors, UIUC / Chinese University of Hong Kong, Shenzhen.
• Fan Wang — person — CoDA co-author, Amazon.
• Runyu Lu, Yubo Wu, Ethan Kou et al. — people — authors of ASPIRE, the strongest LIBERO-Pro baseline CoDA explicitly beats by 7–9 points.
• Xueyang Zhou, Yangming Xu et al. — people — authors of LIBERO-Pro, the external benchmark used to test CoDA's generalization.
• Haoqiang Kang, Xiaokang Ye et al. — people — authors of SimWorld Studio, the strongest closed-loop baseline in CoDA's main four-family experiment.
• Physical Intelligence — organization — makers of π0 and π0.5, the vision-language-action model CoDA uses as its primary policy backbone.
• potential-based reward shaping — concept — the RL technique underlying Text2potential; recurs across robot-learning and RL papers and would benefit from its own note distinguishing it from free-form LLM-synthesized dense rewards.
• vision-language model as failure critic — concept — the specific role a VLM plays in CoDA (diagnosing rollout failures from video/contact/joint evidence, rather than generating actions or scenes directly); worth distinguishing from VLM-as-policy and VLM-as-scene-generator uses elsewhere in the embodied-AI literature.
• LIBERO-Pro — term/concept — external VLA-robustness benchmark testing generalization under position and instruction perturbations, distinct from the original LIBERO benchmark.