Pith. sign in

REVIEW 4 major objections 8 minor 37 references

A robot that has probed an object once can skip the probe forever after by remembering a short, instance-keyed procedure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 15:25 UTC pith:TMEQM3KG

load-bearing objection Solid object-centric efficiency layer with real ablations; the 16–30% op cut is credible, but the abstract overclaims “success never regresses” and real-robot n is thin. the 4 major comments →

arxiv 2607.23702 v1 pith:TMEQM3KG submitted 2026-07-26 cs.RO

Try Once, Then Optimal: De-Redundified Procedure Memory for Cross-Episode Exploration Amortization

Haizhou Ge , Haochen Ouyang , Zhixing Chen , Yufei Jia , Yue Li , Lu Shi , Lei Han , Guyue Zhou
show 1 more author
Ruqi Huang
This is my paper
classification cs.RO
keywords manipulationmemoryexploration amortizationvision-language modelsarticulated objectsinstance-oriented memoryprocedure conditioning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Robots facing objects with hidden internal state—latched microwaves, locked cabinets, twist-or-lift bottle caps—must probe before they can act. Existing memories store successes by state to raise success rates, so the same object still gets re-probed on every later visit. This paper argues that the right unit of memory is the object instance itself: from a single uncovering encounter, distill a short de-redundified procedure, key it on the object’s appearance, and feed it back only as a soft bias on a feedback-driven policy. On four articulated-object tasks (two simulated, two on a real arm), an oracle of that memory cuts operations 16–30% on repeat encounters without hurting success; an off-the-shelf vision-language model recovers most of the saving with no task-specific training. Wrong memories degrade into extra operations, not failures.

Core claim

Instance-Oriented Memory amortizes exploration across encounters with the same object: one uncovering encounter yields a de-redundified procedure keyed on the instance, and later encounters recall it instead of re-exploring. Across microwave, door, bottle, and cabinet, oracle procedure memory cuts operations 16–30% at non-regressing success; a training-free VLM recovers 69–88% of that saving; and incorrect recalls are recovered from rather than obeyed.

What carries the argument

Instance-Oriented Memory (IOM): three modules—procedure distillation from an encounter, instance-keyed memory on identifiable features ϕ(x), and a history-dependent policy conditioned on the recalled procedure as a soft bias—so a hit skips probing and a miss or wrong memory falls back to exploration.

Load-bearing premise

The system assumes a small, hand-built set of candidate procedures fully covers each task, and that objects that need different procedures can be told apart by appearance alone.

What would settle it

On a held-out set of the same articulated tasks, measure operations and success on encounters 2…N: if instance-keyed procedure memory does not cut operations versus re-exploration at matched success, or if wrong recalls drive success down rather than only adding retries, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Repeat encounters with the same physical object become cheaper without retraining the low-level policy.
  • Success-oriented trajectory memories can be layered with object-centric procedure memory rather than replaced by it.
  • An off-the-shelf VLM can serve as both distiller and recognizer, so the memory layer needs no task-specific training.
  • Because the procedure is only a soft bias, imperfect recall costs efficiency, not task completion.
  • The same store can be shared across separately trained single-task policies that face the same instances.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If foundation visuomotor policies already defer to interaction feedback when language and sensation conflict, the dedicated soft-bias training stage may become unnecessary.
  • Widening the instance key beyond appearance to scene context would directly attack the perceptual-aliasing failure mode the authors flag but leave open.
  • A meta-memory that overwrites entries after retries would turn one-shot amortization into a self-correcting loop without changing the three-module design.
  • The same object-centric procedure store could amortize other hidden-state costs (tool configuration, fixture setup) beyond articulated latches and locks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper introduces Instance-Oriented Memory (IOM), an object-centric cross-episode memory framework for manipulation tasks with hidden internal state (e.g., a latched microwave). The idea: a first encounter must probe to uncover the hidden state; IOM distills a short "de-redundified" procedure from that encounter, stores it keyed on the instance's appearance features ϕ(x), and injects it as a soft conditioning bias into a history-dependent, flow-matching policy on later encounters, so redundant probes are skipped. Three modules (distillation, instance-keyed memory, procedure-conditioned policy) are given generic signatures and then instantiated with an off-the-shelf VLM (no task-specific training). On four articulated-object tasks (microwave, door in sim; bottle, cabinet on a real AIRBOT arm), an oracle procedure memory reduces atomic manipulation operations by 16–30% versus a no-memory (Random-procedure) baseline, and the VLM instantiation recovers 69–88% of that saving. Wrong recalls (~12% of door instances) are claimed to be recovered from rather than obeyed, owing to the soft-bias conditioning training of §III-E.

Significance. The problem formulation is clean and, to my knowledge, novel as stated: existing cross-episode memories (RT-Cache, Retrieve-then-Steer, MAP-VLA) key on state and target success, not per-instance operation efficiency, and the paper is explicit and accurate about this distinction. The experimental design has real strengths worth crediting: all memory variants run the same policy from the same initial states, so the Random/GT/VLM comparisons isolate the stored procedure; the efficiency claim is supported mechanistically (the executed-procedure distribution shift on microwave, §IV-B(a)) rather than only by end metrics; and the robustness experiment (wrong VLM procedures on door surfacing as bounded extra operations, not failures) is a falsifiable, informative test of the soft-bias claim. A training-free instantiation that recovers 69–88% of the oracle saving with no task-specific training is practically significant for deployment. The main limitations are statistical: the conjunctive abstract claim ("16–30% reduction AND success never regresses") is only half supported at the reported sample sizes, and the real-robot evaluation is thin (n=10 per condition, apparently one physical insta

major comments (4)
  1. [Abstract / Table III] The abstract states 'success never regresses,' but Table III (microwave row) reports GT success 0.980 versus Random 1.000 — a literal regression for the oracle variant. Over 250 episodes (50 instances × 5) this is roughly 5 failures versus 0. The text's fallback ('stays within two points in simulation,' §IV-B(a)) is not a statistical criterion. Either the claim must be weakened (e.g., 'no statistically significant regression') and backed by an appropriate test (e.g., a one-sided non-inferiority test with a stated margin, or at minimum Wilson confidence intervals per cell), or the abstract must be brought into line with the table. As written, the headline claim is contradicted by the paper's own data.
  2. [§IV, Table III (Real robot rows)] The real-robot success claims rest on n=10 per condition (locked/unlocked × 5 trials, §IV intro). The reported improvements (bottle 0.80→0.90, cabinet 0.70→0.90) are differences of one and two trials respectively; Wilson CIs for 7/10 and 9/10 overlap heavily, and no intervals or tests are given. The efficiency numbers on the real robot are similarly supported by ~8–9 successful episodes per cell. The paper should (i) report confidence intervals for all real-robot cells, (ii) soften 'on the real robot even improves' to a statement compatible with the sample size, and (iii) clarify how many distinct physical instances were used — the protocol description suggests a single bottle and a single cabinet, in which case instance-keyed recognition (a core component) is effectively untested on hardware and the real-robot claims should be scoped accordingly.
  3. [§IV-A(d) Metrics / Table III] Operation count is computed per successful episode while success rates differ across variants, so each variant's mean conditions on a different subset of episodes. The authors argue this is conservative for Random (its failures fall on locked instances, which require longer procedures, §IV-B(a)) — plausible. But the same selection logic is not examined for the other variants, and it can cut the other way: on microwave, GT has the lowest success (0.980) and the lowest ops (1.37); if GT's failures occur disproportionately on expensive episodes, part of its apparent saving is a selection effect. Please either report an unconditional cost metric (expected operations per episode, counting failures at timeout cost or censoring appropriately) or report ops stratified by hidden state, so that the conditioning is symmetric and the 16–30% figures are attributable to the memory rather than to which
  4. [§III-F / Abstract] §III-F constructs all three sets — candidate procedures K, uncovering probes U, and training conditions K◦(p) — by hand from each task's known structure. This is a reasonable isolation choice, but it scopes the 'training-free' and 'out of the box' claims: what is training-free is the distillation and memory modules given a hand-specified procedure set; the amortization claim presupposes that K is complete and that U correctly separates the hidden states. For the four tasks here (binary hidden state, K exhaustively listed in Table I) this is benign, but the abstract and §V should state explicitly that sourcing K (Eq. 5) is outside the evaluated system, and §III-F's proposals (LLM-sourced K, manual retrieval) are unvalidated. Without this scoping, a reader can over-read the deployment readiness of the full pipeline.
minor comments (8)
  1. [Table III / §IV-B(d)] Table III averages over all 5 episodes including the first, memory-free encounter, while §IV-B(d) notes the per-repeat reduction rises to 20–33% when the first encounter is excluded. Since the first-encounter dilution is a modeling choice (one could equally report the steady-state regime), the repeat-only numbers deserve a table column or a supplementary table rather than an inline sentence; they are arguably the more honest headline for 'amortization.'
  2. [§IV-A(c) / §III-B] The VLM is identified only as 'GPT-5.5' (§IV-A(c)). Please report the exact model version/snapshot, temperature, prompt templates (the task prior g and hints of §III-B), and whether distillation and recognition use the same prompt family; these materially affect reproducibility of the 69–88% recovery figure. Relatedly, 'code will be released upon acceptance' — the prompts at minimum should be in the appendix.
  3. [§III-C, Eq. (3) / Table I] Notation: Eq. (3) returns 'u' for a memory miss but the text variously writes u ∈ U and 'a sequence from U'; Table I lists U as a set containing one probe per task. Please make consistent whether U is a set of probes or a singleton, and whether the miss branch draws randomly from U when |U|>1.
  4. [§III-E / Table II] §III-E: the definition of K◦(p) ('conditions that open with p's operations before its first retry but are otherwise unconstrained') is hard to parse on first reading; Table II clarifies it, but one worked sentence in the text (e.g., why [open] rather than [release, open] is the condition for the microwave example) would help. Also state whether K is drawn uniformly from K◦(p) during training and how often the empty/exploratory condition is sampled.
  5. [§IV-B(e)] §IV-B(e) (bottle cap gap 'too fine to read reliably from images') is an interesting ancillary benefit but is asserted without evidence; a single quantitative line (e.g., a VLM or classifier's accuracy at judging cap state from initial frames) would turn it from an anecdote into a result. Alternatively, mark it explicitly as a qualitative observation.
  6. [Fig. 2 / Fig. 3] Fig. 2 is dense: the 'shared spine' and 'first encounter' flows are described in a long caption that carries content absent from the text (e.g., the hit/miss branch). Consider moving the branch logic into §III-C and simplifying the caption. Fig. 3 would benefit from labeling which operation in each sequence is the redundant probe.
  7. [§III-A] The reset vs. persistence distinction (§III-A, final paragraph) is well drawn, but the claim that 'IOM provides the same value in both settings' is asserted, not shown — under persistence the stored procedure can become stale after a successful encounter changes the state, which interacts with the un-implemented meta-memory of §III-C. One sentence acknowledging this interaction (and why reset suffices to demonstrate the core claim) would prevent over-reading.
  8. [References / Acknowledgments] Several references ([2], [11], [20], [28], [29], [32], [33], [36], [37]) are very recent arXiv preprints; please verify venue/status at camera-ready. The acknowledgment of coding-agent assistance is appreciated and appropriately specific.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose efficiency claims come from held-out rollouts, not identities forced by construction.

full rationale

IOM’s central claim is an empirical efficiency result: conditioning a shared procedure-conditioned policy on an oracle (GT) or VLM-recalled de-redundified procedure cuts atomic operations 16–30% versus Random re-exploration on repeat encounters, at non-regressing success. That comparison is produced by actual rollouts under three memory variants on the same policy and initial states (Table III, §IV), not by algebraic rearrangement of fitted inputs. Procedure sets K, U, and training conditions K◦(p) are hand-fixed from each task’s known binary hidden-state structure (§III-F, Tables I–II) expressly to isolate memory effects from sourcing error; they are experimental controls, not parameters fitted to the reported metric and then relabeled as predictions. Soft-bias training (§III-E) teaches the policy to open with K yet recover on feedback—a design choice whose recovery is then measured on wrong recalls (§IV-B), not assumed by definition. Related-work citations (AdaManip, RT-Cache, etc.) supply baselines and policy scaffolding; none is a self-authored uniqueness theorem that forces the efficiency claim. Perceptual-aliasing and incomplete-K limitations are deferred openly (§V), not papered over by circular definition. Honest finding: derivation chain is self-contained experimental comparison; score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 2 invented entities

The central efficiency claim rests on a small set of domain assumptions (hidden state revealable by interaction; finite discrete procedure set; appearance keys instances) plus hand-fixed candidate sets and standard imitation-learning machinery. No new physical entities. Free parameters are ordinary training/demo choices, not constants fitted to produce the reported percentage savings.

free parameters (3)
  • Number of demonstrations (20 sim / 60 real per task) = 20 sim, 60 real
    Chosen by authors following AdaManip configuration; affects policy quality and thus absolute operation counts.
  • Action chunk horizon H and flow-matching training hyperparameters
    Policy architecture knobs adapted from AdaManip/diffusion-to-flow-matching; not swept in the paper for the efficiency claim.
  • Hand-constructed candidate procedure set K and uncovering set U per task = Binary K per task (Table I)
    Fixed exactly from known task structure (§III-F) to isolate memory modules; quality of K directly bounds achievable amortization.
axioms (6)
  • domain assumption An object instance has a hidden internal state s(x) fully revealable by at least one uncovering sequence u ∈ U.
    Stated in §III-A; without full revelation, distillation cannot produce a correct non-redundant procedure.
  • domain assumption Between encounters the hidden state resets (experiments use only the reset setting).
    §III-A explicitly restricts evaluation to reset; persistence would make optimal procedures history-dependent.
  • domain assumption Identifiable features ϕ(x) (appearance) distinguish instances that require different procedures.
    Memory keying assumption; §V acknowledges perceptual aliasing as an open failure mode.
  • domain assumption A history-dependent policy trained with random draws from K◦(p) will follow a correct K and recover via feedback from an incorrect K (soft bias).
    §III-E training design; load-bearing for the non-regression and wrong-memory robustness claims.
  • ad hoc to paper Off-the-shelf VLM can map encounter video+actions plus language prior g to a procedure in K without task-specific training.
    Instantiation assumption (§III-B); residual GT–VLM gap attributed to VLM accuracy (~88% on door).
  • domain assumption Atomic operation counts are the right efficiency metric and timeouts count as failures.
    Metrics §IV-A; standard but choice affects reported percentages.
invented entities (2)
  • Instance-Oriented Memory (IOM) framework no independent evidence
    purpose: Modular object-centric store of de-redundified procedures keyed by ϕ(x) and injected as soft bias.
    Core contribution; three module signatures (distillation D, memory M, policy π). Evaluated empirically; not a physical entity.
  • De-redundified procedure set K and map κ(s(x)) independent evidence
    purpose: Formalize non-redundant instance-level procedures and the cost c(p) of redundant ops.
    Definitional scaffolding in §III-A; standard discrete-procedure abstraction rather than a new ontological claim.

pith-pipeline@v1.2.0-grok45-kimik3 · 16458 in / 3451 out tokens · 61937 ms · 2026-07-30T15:25:20.012207+00:00 · methodology

0 comments
read the original abstract

Manipulating objects with hidden internal state, such as a latched microwave, forces a robot to probe before it can act. Yet a robot that has solved an instance once re-runs the same probes whenever it encounters that instance again, because existing cross-episode memories target task success and organize reuse around states, not the object or the cost of re-exploring it. We present Instance-Oriented Memory (IOM), an object-centric framework that amortizes this exploration: from a single encounter that uncovers the hidden state, whether or not it succeeds, IOM records a short procedure for manipulating that instance, keys it on the object's identifiable features, and injects it as a soft bias on a procedure-conditioned policy. A later encounter recognizes the object and recalls its procedure instead of re-exploring. We instantiate this distillation with an off-the-shelf vision-language model (VLM) that parses each encounter into the procedure without task-specific training. Across four articulated-object tasks, two in simulation (microwave, door) and two on a real robot (bottle, cabinet), an oracle procedure memory cuts manipulation operations by 16-30% over re-exploration at non-regressing success, and the VLM instantiation recovers 69-88% of that saving out of the box. Because the procedure is a soft bias on a feedback-driven policy, an incorrect memory is recovered from rather than obeyed: success holds even when a retrieved procedure is wrong, as for $\approx$12% of door instances. Across all tasks the benefit is purely one of efficiency: success never regresses, and on the real robot even improves. Code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2607.23702 by Guyue Zhou, Haizhou Ge, Haochen Ouyang, Lei Han, Lu Shi, Ruqi Huang, Yue Li, Yufei Jia, Zhixing Chen.

Figure 1
Figure 1. Figure 1: Instance-Oriented Memory (IOM). An object with hidden internal state (here a latched microwave) must be probed before it can be manipulated, and that probe is pure overhead on every later encounter. IOM removes it. Top: the first encounter probes, which uncovers the hidden state whether or not the attempt succeeds. Middle: distillation strips the redundant probe into a short de-redundified procedure, writt… view at source ↗
Figure 2
Figure 2. Figure 2: The Instance-Oriented Memory (IOM) framework. IOM amortizes exploration across repeated encounters with the same object instance through three modules: procedure distillation, instance-keyed memory, and a procedure-conditioned policy. Bottom (shared spine): an object instance is keyed by its identifiable features ϕ(x) into the instance-keyed memory; a hit recalls the stored procedure Kˆ , a miss returns an… view at source ↗
Figure 3
Figure 3. Figure 3: Experimental tasks. We evaluate on four articulated-object tasks with hidden internal state, two in simulation (door, microwave) and two on a real robot (bottle, cabinet). For each task we show the initial observation (leftmost) followed by the ordered atomic operations that compose its manipulation procedure (left to right); the hidden state (a latch or lock, which way a handle must turn, or whether a cap… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 30 linked inside Pith

  1. [1]

    Rt-cache: Training-free retrieval for real-time manipulation,

    O. Kwon, A. George, A. Bartsch, and A. Barati Farimani, “Rt-cache: Training-free retrieval for real-time manipulation,” inIEEE-RAS Intl. Conf. on Humanoid Robots (Humanoids), 2025, arXiv:2505.09040

    Pith/arXiv arXiv 2025

  2. [2]

    Retrieve-then-steer: Online success memory for test-time adaptation of generative vlas,

    J. Zhao, H. Yang, Y . Hu, Y . Gao, Q. Ou, C. Wan, S. Dong, Z. Ma, and Y . Gong, “Retrieve-then-steer: Online success memory for test-time adaptation of generative vlas,”arXiv preprint arXiv:2605.10094, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Adamanip: Adaptive articulated object manipulation environments and policy learning,

    Y . Wanget al., “Adamanip: Adaptive articulated object manipulation environments and policy learning,” inInternational Conference on Learning Representations (ICLR), 2025, arXiv:2502.11124

    Pith/arXiv arXiv 2025

  4. [4]

    Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions,

    ——, “Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions,” inEuropean Conference on Computer Vision (ECCV), 2022, arXiv:2112.00246

    Pith/arXiv arXiv 2022

  5. [5]

    Map-vla: Memory-augmented prompting for vision-language-action model in robotic manipulation,

    R. Liet al., “Map-vla: Memory-augmented prompting for vision-language-action model in robotic manipulation,” 2025, arXiv:2511.09516

    arXiv 2025

  6. [6]

    Bring my cup! personalizing vision-language-action models with visual attentive prompting,

    S. Lee, S. Mo, and W.-S. Han, “Bring my cup! personalizing vision-language-action models with visual attentive prompting,” 2025, arXiv:2512.20014

    Pith/arXiv arXiv 2025

  7. [7]

    Memer: Scaling up memory for robot control via experience retrieval,

    A. Sridhar, J. Pan, S. Sharma, and C. Finn, “Memer: Scaling up memory for robot control via experience retrieval,” 2026, arXiv:2510.20328

    arXiv 2026

  8. [8]

    Online object model reconstruction and reuse for lifelong improvement of robot manipulation,

    S. Lu, R. Wang, Y . Miao, C. Mitash, and K. Bekris, “Online object model reconstruction and reuse for lifelong improvement of robot manipulation,” 2021, arXiv:2109.13910

    Pith/arXiv arXiv 2021

  9. [9]

    Adaptive articulated object manipulation on the fly with foundation model reasoning and part grounding,

    X. Zhanget al., “Adaptive articulated object manipulation on the fly with foundation model reasoning and part grounding,” 2025, arXiv:2507.18276

    Pith/arXiv arXiv 2025

  10. [10]

    Vat-mart: Learning visual action trajectory proposals for manipulating 3d articulated objects,

    R. Wuet al., “Vat-mart: Learning visual action trajectory proposals for manipulating 3d articulated objects,” inInternational Conference on Learning Representations (ICLR), 2022, arXiv:2106.14440

    Pith/arXiv arXiv 2022

  11. [11]

    Beyond short-horizon: Vq-memory for robust long- horizon manipulation in non-markovian simulation benchmarks,

    H. Wanget al., “Beyond short-horizon: Vq-memory for robust long- horizon manipulation in non-markovian simulation benchmarks,” 2026, arXiv:2603.09513

    arXiv 2026

  12. [12]

    Strap: Robot sub-trajectory retrieval for augmented policy learning,

    M. Memmel, J. Berg, B. Chen, A. Gupta, and J. Francis, “Strap: Robot sub-trajectory retrieval for augmented policy learning,” inIn- ternational Conference on Learning Representations (ICLR), 2025, arXiv:2412.15182

    Pith/arXiv arXiv 2025

  13. [13]

    Learning and retrieval from prior data for skill-based imitation learning,

    S. Nasiriany, T. Gao, A. Mandlekar, and Y . Zhu, “Learning and retrieval from prior data for skill-based imitation learning,” inConference on Robot Learning (CoRL), 2022, arXiv:2210.11435

    Pith/arXiv arXiv 2022

  14. [14]

    Keypoint action tokens enable in-context imitation learning in robotics,

    N. Di Palo and E. Johns, “Keypoint action tokens enable in-context imitation learning in robotics,” inRobotics: Science and Systems (RSS), 2024, arXiv:2403.19578

    Pith/arXiv arXiv 2024

  15. [15]

    In-context imitation learning via next-token prediction,

    L. Fuet al., “In-context imitation learning via next-token prediction,” 2024, arXiv:2408.15980

    Pith/arXiv arXiv 2024

  16. [16]

    Instant policy: In-context imitation learning via graph diffusion,

    V . V osylius and E. Johns, “Instant policy: In-context imitation learning via graph diffusion,” inInternational Conference on Learning Repre- sentations (ICLR), 2025, arXiv:2411.12633

    Pith/arXiv arXiv 2025

  17. [17]

    Reflexion: Language agents with verbal reinforcement learning,

    N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,” inConference on Neural Information Processing Systems (NeurIPS), 2023, arXiv:2303.11366

    Pith/arXiv arXiv 2023

  18. [18]

    Expel: Llm agents are experiential learners,

    A. Zhao, D. Huang, Q. Xu, M. Lin, Y .-J. Liu, and G. Huang, “Expel: Llm agents are experiential learners,” inAAAI Conference on Artificial Intelligence (AAAI), 2024, arXiv:2308.10144

    Pith/arXiv arXiv 2024

  19. [19]

    Generative agents: Interactive simulacra of human behavior,

    J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” inACM Symposium on User Interface Software and Technology (UIST), 2023, arXiv:2304.03442

    Pith/arXiv arXiv 2023

  20. [20]

    Experiential reflec- tive learning for self-improving llm agents,

    M.-A. Allard, A. Teinturier, V . Xing, and G. Viaud, “Experiential reflec- tive learning for self-improving llm agents,” 2026, arXiv:2603.24639

    arXiv 2026

  21. [21]

    Position: Episodic memory is the missing piece for long-term llm agents,

    M. Pinket al., “Position: Episodic memory is the missing piece for long-term llm agents,” 2025, arXiv:2502.06975

    Pith/arXiv arXiv 2025

  22. [22]

    Distilling and retrieving generalizable knowledge for robot manipulation via language corrections,

    L. Zhaet al., “Distilling and retrieving generalizable knowledge for robot manipulation via language corrections,” 2023, arXiv:2311.10678

    Pith/arXiv arXiv 2023

  23. [23]

    Lifelong robot library learning: Bootstrap- ping composable and generalizable skills for embodied control with language models,

    G. Tziafas and H. Kasaei, “Lifelong robot library learning: Bootstrap- ping composable and generalizable skills for embodied control with language models,” 2024, arXiv:2406.18746

    Pith/arXiv arXiv 2024

  24. [24]

    Open-ended in- structable embodied agents with memory-augmented large language models,

    G. Sarch, Y . Wu, M. J. Tarr, and K. Fragkiadaki, “Open-ended in- structable embodied agents with memory-augmented large language models,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2023, arXiv:2310.15127

    Pith/arXiv arXiv 2023

  25. [25]

    V oyager: An open-ended embodied agent with large language models,

    G. Wanget al., “V oyager: An open-ended embodied agent with large language models,” 2023, arXiv:2305.16291

    Pith/arXiv arXiv 2023

  26. [26]

    Reflect: Summarizing robot experiences for failure explanation and correction,

    Z. Liu, A. Bahety, and S. Song, “Reflect: Summarizing robot experiences for failure explanation and correction,” inConference on Robot Learning (CoRL), 2023, arXiv:2306.15724

    Pith/arXiv arXiv 2023

  27. [27]

    Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,

    H. Shiet al., “Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,” 2025, arXiv:2508.19236

    Pith/arXiv arXiv 2025

  28. [28]

    Eventvla: Event-driven visual evidence memory for long-horizon vision-language-action policies,

    G. Yanget al., “Eventvla: Event-driven visual evidence memory for long-horizon vision-language-action policies,” 2026, arXiv:2606.20092

    Pith/arXiv arXiv 2026

  29. [29]

    Mem: Multi-scale embodied memory for vision language action models,

    M. Torneet al., “Mem: Multi-scale embodied memory for vision language action models,” 2026, arXiv:2603.03596

    arXiv 2026

  30. [30]

    Robocat: A self-improving generalist agent for robotic manipulation,

    K. Bousmaliset al., “Robocat: A self-improving generalist agent for robotic manipulation,” 2024, arXiv:2306.11706

    Pith/arXiv arXiv 2024

  31. [31]

    Self-improving robots: End-to-end autonomous visuomotor reinforcement learning,

    A. Sharma, A. M. Ahmed, R. Ahmad, and C. Finn, “Self-improving robots: End-to-end autonomous visuomotor reinforcement learning,” in Robotics: Science and Systems (RSS), 2023, arXiv:2303.01488

    Pith/arXiv arXiv 2023

  32. [32]

    Robomemarena: A comprehensive and challenging robotic memory benchmark,

    H. Leiet al., “Robomemarena: A comprehensive and challenging robotic memory benchmark,” 2026, arXiv:2605.10921

    Pith/arXiv arXiv 2026

  33. [33]

    Meta-memory for large language models,

    B. Liang, C. Ke, R. Zhao, Q. Zhu, L. Gui, Y . Yu, H. Wang, R. Xu, and K.-F. Wong, “Meta-memory for large language models,”IEEE Transactions on Audio, Speech and Language Processing, vol. 34, pp. 2774–2787, 2026

    2026

  34. [34]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inRobotics: Science and Systems (RSS), 2023, arXiv:2304.13705

    Pith/arXiv arXiv 2023

  35. [35]

    M3e: Moka massive mixed embedding model,

    Y . Wang, Q. Sun, and S. He, “M3e: Moka massive mixed embedding model,” https://huggingface.co/moka-ai/m3e-small, 2023

    2023

  36. [36]

    LA4VLA: Learning to act without seeing via language-action pretraining,

    T. Lin, Y . Du, Y . Mao, Z. Ye, Y . Zhong, B. Cheng, Y . Wang, J. Liu, Y . Tian, J. Yanet al., “LA4VLA: Learning to act without seeing via language-action pretraining,” 2026, arXiv:2606.27295

    Pith/arXiv arXiv 2026

  37. [37]

    Chameleon: Control-indexed prospective memory for visuomotor manipulation,

    X. Guoet al., “Chameleon: Control-indexed prospective memory for visuomotor manipulation,” 2026, arXiv:2603.24576

    Pith/arXiv arXiv 2026