Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Legacy demonstration data from older hardware helps an upgraded robot only after the new configuration reaches a task-dependent competence threshold, with gains rising sharply past the threshold before diminishing near saturation.

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 · deepseek-v4-flash

2026-08-01 01:58 UTC pith:COTSXW3S

load-bearing objection A practically useful three-phase claim about when legacy robot data starts to help, with real held-out support but a confounded flagship transition and a partly circular threshold definition. the 5 major comments →

arxiv 2607.25593 v1 pith:COTSXW3S submitted 2026-07-28 cs.RO

When Does Legacy Data Start to Help? Emergent Transfer in Cross-Configuration Robot Learning

Tao Wang , Hudson Hou , Yingdong Hu , Yufeng Liu , Qinghai Li , Yingjie Jiang , Yingzhi Wang , Cheng Ma
show 2 more authors
Richard Wang Yang Gao
This is my paper
classification cs.RO
keywords robot learningcross-configuration transferlegacy demonstration dataco-trainingtransfer thresholdinverted-U gainbehavior cloninghardware iteration
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.

This paper asks when previously collected demonstration data from an older robot configuration begins to benefit a policy for upgraded hardware, after cameras and grippers change. The authors claim that the gain from co-training old and new data is not monotonic: it stays at zero until the new-hardware policy reaches a task-dependent transfer threshold, then jumps sharply before diminishing near saturation. They establish this three-phase pattern across real-robot insertion and grasping tasks, with flower insertion improving from 23.3% to 86.7% once the threshold is crossed, while the same co-training leaves a 10% baseline unchanged. They also provide a theoretical mechanism based on gradient alignment and residual policy uncertainty, plus a phase-aware data collection rule that saved 6.5 hours of new data on a held-out mobile watering task. If correct, the result changes when engineers should merge legacy data: not immediately, only after the new robot has learned enough task structure.

Core claim

The central discovery is an emergent transition in cross-configuration co-training: the benefit of adding legacy demonstrations to new-hardware training is governed by the upgraded configuration's standalone success rate. For any task there is a transfer threshold τ(T); below it, co-training yields no measurable gain, above it the gain rises sharply, then declines near saturation. The supporting evidence is the flower-insertion contrast (10.0%→10.0% below threshold vs 23.3%→86.7% above) and the pen-insertion saturation (85.0%→93.3%). The authors model the phenomenon with a stage-decodability order parameter: once the policy's internal representation can reliably separate task stages, legacy

What carries the argument

The transfer threshold τ(T), defined as the smallest standalone success rate above which co-training with legacy data yields positive expected gain. It is coupled to a stage-decodability order parameter ρ(T;θ): the probability that the policy's internal representation can be decoded into the task's latent stages (approach, grasp, insert). The argument runs through two theorems: a threshold transition in expected gradient alignment ⟨g_S, g_T⟩ (non-positive below τ, positive above) and an inverted-U gain law E[ΔSR|SR] = [κ(1−SR)−δ(SR)]·1[SR>τ(T)]. The stage-decodability proxy connects the unobservable representation to the observable success rate, and residual uncertainty η(1−SR) drives the de

Load-bearing premise

The sharp transition claim rests on assuming that the new robot's standalone success rate reliably tracks an unmeasured internal quantity — how cleanly its representation separates task stages — so that the threshold is real rather than a side effect of how the two flower-insertion datasets were collected.

What would settle it

Collect new-hardware demonstration batches for the same Gen-2 flower-insertion task at several intermediate budgets between the 4.3h (10.0% baseline) and 15.6h (23.3% baseline) points, co-train each with the same legacy set, and plot ΔSR against baseline; a smooth upward slope rather than a near-zero-then-jump curve would falsify the phase transition. Independently, recording per-source gradient alignment during such co-training would test whether expected alignment is actually non-positive below the threshold.

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

If this is right

  • Engineers should not blindly mix legacy and new-hardware data: below the threshold, legacy data yields nothing, so collection effort should go to raising the new configuration's standalone success.
  • Once the threshold is crossed, co-training can deliver large gains, with flower insertion improving by 63.4 percentage points (23.3%→86.7%).
  • Data quality affects how quickly a task crosses phases but does not bypass the threshold: refined pen-insertion data reached a higher baseline with fewer hours, and the phase then predicted the co-training gain.
  • The phase-aware rule cut required new-hardware collection time from 8 hours to 1.5 hours on the held-out watering task, where co-training at 1.5h produced gains of 38–40 points.
  • The model predicts that τ(T) grows with task complexity H(T)=L(T)log(1/ε(T)), so longer or tighter-tolerance tasks need higher standalone baselines before legacy data helps.

Where Pith is reading between the lines

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

  • If the phase transition is a general property of cross-domain co-training rather than a quirk of this vision-language-action setup, similar inverted-U gains should appear under other embodiment or dataset shifts (e.g., sim-to-real, new action spaces), making the transfer threshold a candidate design principle for generalist robot datasets.
  • A testable extension: measure per-source gradient alignment during co-training (which the paper did not record) to verify that alignment actually flips sign at τ(T); if alignment stays positive below the threshold, the mechanism would need replacement even if the empirical pattern holds.
  • The inverted-U suggests an optimal data-mixing schedule: instead of co-training from the start, one could first train on new-hardware data alone until the baseline crosses τ(T), then introduce legacy data — a curriculum for data mixing.
  • The threshold's dependence on task complexity H(T) implies that easy, coarse-tolerant tasks (like watering) become legacy-ready much sooner than precise insertion tasks, enabling task-level prioritization of scarce new-hardware collection effort.

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

5 major / 5 minor

Summary. The paper studies cross-configuration robot learning on a wheeled humanoid platform with two hardware generations. It compares policies trained only on new-hardware (Gen-2) demonstrations with policies co-trained on legacy Gen-1 plus Gen-2 data, and examines how the co-training gain ΔSR = SR_co-train − SR_single varies with the standalone success rate SR. The authors report a three-phase pattern: no measurable gain at low SR, large gains at intermediate SR, and diminishing returns near saturation. They introduce a task-dependent transfer threshold τ(T) and an inverted-U gain law motivated by gradient alignment and residual policy uncertainty, and propose a phase-aware data collection rule: collect new-hardware data until standalone competence crosses the threshold, then co-train with legacy data. Validation on a held-out mobile dual-arm watering task used three new-data budgets (0.5/1.5/8 h) and observed no gain at 0.5 h, +38–40 percentage-point gains at 1.5 h, and no significant gain at 8 h. The paper reports Fisher exact tests and Wilson intervals for all key comparisons.

Significance. If the three-phase pattern holds, it has immediate practical value for hardware iteration: rather than always mixing legacy data, a practitioner should first bring the new configuration to a minimal competence level. The held-out watering experiment is a genuine strength because it varies only the new-hardware data budget while holding legacy data, sampling ratio, optimization, and evaluation fixed, and the statistical reporting is solid. However, the central claim of an emergent, threshold-like transition is not yet robustly established: the flagship flower-insertion contrast is confounded by dataset quantity and quality, the threshold is defined from the same sign-change it is used to explain, and the theory relies on unmeasured assumptions. The paper nevertheless makes a useful empirical observation that co-training benefit is non-monotonic in standalone competence, and it demonstrates a practically relevant data-allocation rule.

major comments (5)
  1. [§4.1–4.2, Tables 1–3] The Phase I→II evidence for flower insertion is confounded. Phase I uses 4.31 h of early Gen-2 data; Phase II uses a separately collected 15.60 h quality-refined batch with stricter operator training, per-trajectory quality filtering, and tighter pose tolerances (§3). The gain jump from 0 to +63.4 percentage points could be caused by the target dataset being larger and cleaner, not by the target policy crossing a competence threshold. To support a sharp threshold, the authors should vary the amount of new-hardware data within a single collection protocol (or otherwise control for dataset quality and size) and include intermediate baselines.
  2. [§5, Definition 1 and Eq. (5)] The threshold τ(T) in Eq. (2) is defined as inf{SR : E[ΔSR|SR] > 0}, and it then appears as the switch in the inverted-U law of Eq. (5). Consequently, the existence of a sign change above τ is true by construction; a piecewise fit of the form [κ(1−SR)−δ(SR)]·1[SR>τ] will trivially describe any data with a positive-gain region unless τ is predicted independently. The predictive content must come from the complexity–threshold relation τ ≈ α + βH(T), but §C.5 admits that 'these observations are consistent with the proposed monotonic relation but are insufficient to establish a precise scaling law.' The paper should distinguish the fitted description from an independent prediction.
  3. [§5, Assumptions 1 and 4; Appendix C.4] The theoretical account rests on Assumption 1 (stage decodability ρ_c(T;θ) is monotonically coupled to standalone SR) and Assumption 4 (residual uncertainty H_within = η(1−SR)). Neither is measured. Appendix C.4 concedes that 'direct per-source gradient alignment was not recorded,' so the loss trajectories are only an indirect diagnostic. The statement that gradient alignment 'explains' the phase structure is therefore too strong; it is a plausible post-hoc model. Please either measure the assumed quantities (e.g., probe stage decodability or estimate remaining uncertainty) or explicitly present the theory as a qualitative framework with no direct empirical support.
  4. [§6.1, Table 5] The held-out watering sweep has only three new-data budgets (0.5/1.5/8 h), with no points between 0.5 and 1.5 h or just above the predicted threshold. A smooth inverted-U function with gradual onset would produce the same three observations; the data cannot distinguish a sharp threshold from a gradual improvement, which is the central claim. Moreover, the positive-gain phase is represented by a single budget (1.5 h), albeit with two sub-stages. Please add an intermediate budget (e.g., 1.0 or 2–3 h) and ideally repeat the sweep to establish reliability.
  5. [Appendix B.2, Table 7] The Phase III saturation comparison for pen insertion is also confounded. The 'Early + refined' condition (32.21 h) is compared to the 'Refined' condition (13.58 h); the change in gain from +26.6 to +8.3 could reflect the larger, mixed dataset rather than the higher standalone success rate alone. Since this is the main evidence for diminishing returns in Phase III, the data-composition difference should be controlled (e.g., use the same union dataset with random subsets of differing size, or compare same-hours refined-only conditions).
minor comments (5)
  1. [Figure 3] The '±22% envelope' is not defined. Is it a confidence band, a model-fit residual, or an ad hoc uncertainty range? Please specify how it was computed.
  2. [Abstract, Introduction] The term 'grokking-like' is potentially misleading: grokking is a training-time phenomenon, whereas here the transition is across data-budget/standalone-competence levels. Consider using a more neutral term such as 'regime-dependent transfer' or clarify the analogy.
  3. [Appendix C] After Assumption 5, the text contains garbled strings like '/uni00000013/...'. This appears to be a PDF-extraction artifact. Please ensure the submitted PDF renders the proofs cleanly.
  4. [Table 4] The phase boundaries are listed as '<15–20%', '20–75%', and '>75%', but Table 3 reports a Phase II result at 71.7% and Phase III at 85%. A fuzzy boundary is fine, but it should be stated as approximate rather than categorical.
  5. [Eq. (4)] The symbol ε_dom is described as 'irreducible configuration-specific mismatch' but is not formally defined in terms of the data distributions or gradients. Please clarify its meaning and units.

Circularity Check

1 steps flagged

Partial circularity: τ is defined as the empirical sign-change point and then reused as the switch in the inverted-U law; the held-out watering prediction is independent.

specific steps
  1. self definitional [Definition 1 (Eq. 2), Section 5; Theorem 2 (Eq. 5), Section 5]
    "Definition 1 (Task-dependent transfer threshold). For task T, the transfer threshold τ(T) is the smallest standalone success rate above which co-training with legacy data yields positive expected gain: τ(T) = inf{SR :E[∆SR|SR]>0}. (2) ... E[∆SR|SR] = [κ(1−SR)−δ(SR)] ×⊮[SR> τ(T)], (5)"

    The switch in the predicted gain law is literally the definition of τ. Eq. (2) defines τ as the infimum of the set of SR values with positive expected co-training gain, and Eq. (5) uses exactly that τ as the indicator cutoff. Therefore the claim that gains are zero below τ and positive above τ is true by construction for any dataset with a sign change; it is an empirical summary of the ΔSR data rather than a derived prediction. The circularity is partial because the held-out watering validation estimates τ from H(T)≈44 and a pilot standalone learning curve, not from ΔSR, providing independent out-of-sample support.

full rationale

The main circularity is definitional: τ(T) is introduced as the sign-change location of E[ΔSR|SR], then the inverted-U law of Eq. (5) uses that same τ as its phase switch. Consequently, the 'prediction' that legacy data is ineffective below τ and beneficial above it cannot fail on the data used to define τ. This does not reduce the entire paper to a tautology, because the watering experiment is genuinely held out: the crossing is estimated from H(T)≈44 and a pilot standalone learning curve, and the 0.5/1.5/8 h sweep then shows the predicted phase pattern. No load-bearing self-citation is present; the cited π0.5 and X-VLA works are external and not used to force the conclusion. The sharper threats to the empirical transition claim—the confounded flower-insertion comparison (4.31 h early vs 15.60 h refined Gen-2 data) and the coarse three-budget watering sweep—are validity/confound concerns rather than circularity. The paper itself concedes in Appendix C.4 that per-source gradient alignment was never recorded and in Section 7 that τ(T)≈α+βH(T) is not a validated scaling law; these limitations weaken the theoretical account but do not turn it into a circular derivation. Score 4 reflects one definitional loop with substantial independent content remaining.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 2 invented entities

The empirical three-phase pattern is the genuinely new content; the theory that explains it adds several free functions and unmeasured latent constructs (τ(T), κ, η, δ(SR), stage decodability). The held-out watering task provides one independent check of the threshold idea, but the core scaling law τ≈α+βH is explicitly labeled an unvalidated prediction by the authors. The central claim therefore rests on a modest set of real-robot measurements plus a scaffolding of assumptions.

free parameters (6)
  • task-dependent transfer threshold τ(T) = ≈15–20% standalone SR for insertion tasks; inferred from observed sign change
    Defined in Definition 1 as inf{SR:E[ΔSR|SR]>0}; its value is read off the measured data rather than derived from task geometry. The paper's phase-aware rule depends on estimating this threshold.
  • co-training gain coefficient κ(T) = not estimated
    Task-dependent constant in Eq. (5) that absorbs stage-routing probability and domain constants; no independent measurement or prediction is given for its value.
  • entropy-coupling coefficient η = not estimated
    Assumption 4 postulates H_within = η(1−SR) to first order; η is a free proportionality constant that converts the entropy ansatz into the (1−SR) decay term.
  • configuration-conflict cost δ(SR) = unspecified function
    Assumption 5 introduces δ(SR) as a non-negative cost that 'may increase near saturation'; its form is free, so Eq. (5) can accommodate the observed diminishing returns without a concrete mechanism.
  • phase boundaries (<15–20%, 20–75%, >75%) = 15–20%, 20–75%, >75%
    Table 4 partitions co-training gain into Phase I/II/III; these boundaries are set post hoc from the points in Figure 3, not predicted independently.
  • complexity-threshold relation τ(T) ≈ α + βH(T) = α, β not stated
    Proposed in Section 6.1 and Appendix C.5; the authors themselves describe it as 'a theoretical prediction supported by initial evidence, not a fully validated scaling law.'
axioms (7)
  • domain assumption Monotone coupling (Assumption 1): stage decodability ρ(T;θ_single) is non-decreasing in standalone success rate SR.
    This is needed to use SR as an observable proxy for latent stage decodability and to map ρ_crit to τ(T). It is plausible but never measured in the paper.
  • domain assumption Shared stage-conditional objective (Assumption 2): legacy and target demonstrations share a per-stage objective up to configuration-specific residual error ε_dom.
    Core premise of Theorem 1: it makes same-stage legacy and target gradients non-negative. No direct evidence is provided that the two hardware generations share stage-conditional optima.
  • standard math Continuity of expected gradient alignment (Assumption 3).
    Standard analytic assumption used to guarantee a crossing point in alignment; not a domain claim.
  • ad hoc to paper Linear capability-entropy coupling (Assumption 4): H_within = η(1−SR) to first order.
    This ansatz is what manufactures the (1−SR) decay in the inverted-U law; it is a modeling choice, not an independently established scaling.
  • ad hoc to paper Bounded saturation interference (Assumption 5): δ(SR) is non-negative, small away from saturation, and may increase as SR→1.
    Introduced to explain why gains diminish near saturation without specifying a concrete conflict mechanism.
  • domain assumption Latent task stages Z(T) exist, are shared across the two hardware generations, and ground-truth stage labels z(s,o) are well-defined.
    The whole stage-decodability formalism (Eqs. 1, 8, 11) depends on this decomposition, but no stage labels are collected or measured in the experiments.
  • domain assumption Fine-tuning from the pretrained π0.5 VLA backbone is an appropriate common initialization for both hardware generations.
    All policies start from this backbone. If the backbone already bridges much of the camera/gripper shift, the observed threshold could be an artifact of that particular pretraining rather than a general law.
invented entities (2)
  • Latent task stages z(s,o) no independent evidence
    purpose: Theoretical mechanism: trajectories decompose into stages (approach, grasp, insert) that control when legacy gradients align with target gradients.
    The stages are never labeled, measured, or validated; they are a postulated hidden variable used to construct the threshold theory.
  • Stage decodability ρ_c(T;θ) no independent evidence
    purpose: Order parameter that supposedly controls gradient alignment and the transfer threshold.
    Defined via an arbitrary decoder g in Eq. (1), but never estimated from the policy's representations. Its monotone coupling to SR is assumed, not shown.

pith-pipeline@v1.3.0-alltime-deepseek · 17716 in / 13021 out tokens · 122052 ms · 2026-08-01T01:58:20.559864+00:00 · methodology

0 comments
read the original abstract

Robotic hardware evolves over time, but demonstration data is often tied to a specific sensor and actuator configuration. This raises a practical and underexplored question: when does legacy data begin to benefit an upgraded robot? We study this question on a wheeled humanoid platform across two hardware generations, where both the camera and gripper are changed while the overall morphology remains fixed. Contrary to the common assumption that more cross-configuration data is always helpful, we observe a grokking-like transition: legacy data remains ineffective until the upgraded configuration acquires a minimum level of task competence, after which co-training gains rise sharply before diminishing near saturation. We hypothesize that this task-dependent transition is governed by a transfer threshold and characterize the resulting three-phase pattern. Across real-robot manipulation tasks, we observe all three phases: no measurable benefit at low competence ($10.0\% \rightarrow 10.0\%$), a sharp gain after crossing the threshold ($23.3\% \rightarrow 86.7\%$ on flower insertion), and diminishing returns at high competence ($85.0\% \rightarrow 93.3\%$ on pen insertion). We provide a theoretical account based on gradient alignment and residual policy uncertainty, and derive a phase-aware rule for deciding when to collect more new-hardware data and when to reuse legacy demonstrations. We further validate this three-phase pattern on a mobile dual-arm watering task, with results consistent with our predictions.

Figures

Figures reproduced from arXiv: 2607.25593 by Cheng Ma, Hudson Hou, Qinghai Li, Richard Wang, Tao Wang, Yang Gao, Yingdong Hu, Yingjie Jiang, Yingzhi Wang, Yufeng Liu.

Figure 1
Figure 1. Figure 1: The three-phase pattern of cross-configuration co-training. We compare policies trained on legacy data [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Robot platform, hardware iteration, and tabletop manipulation tasks. (a) The robot comprises dual 7-DoF arms, a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Phase transition in cross-configuration co-training. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mobile dual-arm watering task used for held-out [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training-loss dynamics across four matched [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation

    cs.RO 2026-07 conditional novelty 6.0

    Action-conditioned world-model verification with conformal first-intervention control and latency-aware suffix repair raises RoboCasa365 success 8.5 points over invocation-matched periodic replanning.

Reference graph

Works this paper leans on

45 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Transactions on Machine Learning Research , year=

    A Generalist Agent , author=. Transactions on Machine Learning Research , year=

  2. [2]

    Nair, Suraj and Rajeswaran, Aravind and Kumar, Vikash and Finn, Chelsea and Gupta, Abhinav , booktitle=

  3. [3]

    2212.06817 , archivePrefix=

    Brohan, Anthony and Brown, Noah and Carbajal, Justice and Chebotar, Yevgen and Dabis, Joseph and Finn, Chelsea and Gopalakrishnan, Keerthana and Hausman, Karol and Herzog, Alexander and Hsu, Jasmine and others , year=. 2212.06817 , archivePrefix=

    Pith/arXiv arXiv

  4. [4]

    Brohan, Anthony and Brown, Noah and Carbajal, Justice and Chebotar, Yevgen and Chen, Xi and Choromanski, Krzysztof and Ding, Tianli and Driess, Danny and Dubey, Avinava and Finn, Chelsea and others , booktitle=

  5. [5]

    IEEE International Conference on Robotics and Automation (ICRA) , year=

    Open X-Embodiment: Robotic Learning Datasets and. IEEE International Conference on Robotics and Automation (ICRA) , year=

  6. [6]

    Robotics: Science and Systems (RSS) , year=

  7. [7]

    Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and Xiao, Ted and Balakrishna, Ashwin and Nair, Suraj and Rafailov, Rafael and Foster, Ethan and Lam, Grace and Sanketi, Pannag and others , booktitle=

  8. [8]

    2510.10274 , archivePrefix=

    Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and Liu, Dongxiu and Kang, Xirui and Feng, Yuchun and Zheng, Yinan and Zou, Jiayin and Chen, Yilun and Zeng, Jia and Zhang, Ya-Qin and Pang, Jiangmiao and Liu, Jingjing and Wang, Tai and Zhan, Xianyuan , year=. 2510.10274 , archivePrefix=

    Pith/arXiv arXiv

  9. [9]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-Trained Transformers , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  10. [10]

    Robotics and Autonomous Systems , volume=

    A Survey of Robot Learning from Demonstration , author=. Robotics and Autonomous Systems , volume=. 2009 , publisher=

    2009

  11. [11]

    International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , author=. International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=

  12. [12]

    Advances in Neural Information Processing Systems (NeurIPS) , pages=

    Generative Adversarial Imitation Learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , pages=

  13. [13]

    Conference on Robot Learning (CoRL) , year=

    What Matters in Learning from Offline Human Demonstrations for Robot Manipulation , author=. Conference on Robot Learning (CoRL) , year=

  14. [14]

    Behavior Transformers: Cloning

    Shafiullah, Nur Muhammad Mahi and Cui, Zichen Jeff and Altanzaya, Ariuntuya and Pinto, Lerrel , booktitle=. Behavior Transformers: Cloning

  15. [15]

    Robotics: Science and Systems (RSS) , year=

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author=. Robotics: Science and Systems (RSS) , year=

  16. [16]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Data Quality in Imitation Learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  17. [17]

    Dasari, Sudeep and Ebert, Frederik and Tian, Stephen and Nair, Suraj and Bucher, Bernadette and Schmeckpeper, Karl and Singh, Siddharth and Levine, Sergey and Finn, Chelsea , booktitle=

  18. [18]

    Ebert, Frederik and Yang, Yanlai and Schmeckpeper, Karl and Bucher, Bernadette and Georgakis, Georgios and Daniilidis, Kostas and Finn, Chelsea and Levine, Sergey , booktitle=

  19. [19]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    A Survey on Transfer Learning , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2010 , publisher=

    2010

  20. [20]

    Machine Learning , volume=

    A Theory of Learning from Different Domains , author=. Machine Learning , volume=. 2010 , publisher=

    2010

  21. [21]

    NIPS Workshop on Inductive Transfer: 10 Years Later , year=

    To Transfer or Not to Transfer , author=. NIPS Workshop on Inductive Transfer: 10 Years Later , year=

  22. [22]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Characterizing and Avoiding Negative Transfer , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  23. [23]

    IEEE/CAA Journal of Automatica Sinica , volume=

    A Survey on Negative Transfer , author=. IEEE/CAA Journal of Automatica Sinica , volume=

  24. [24]

    Proceedings of the National Academy of Sciences , volume=

    Overcoming Catastrophic Forgetting in Neural Networks , author=. Proceedings of the National Academy of Sciences , volume=

  25. [25]

    Neural Networks , volume=

    Continual Lifelong Learning with Neural Networks: A Review , author=. Neural Networks , volume=. 2019 , publisher=

    2019

  26. [26]

    Wan, Weikang and Zhu, Yifeng and Shah, Rutav and Zhu, Yuke , booktitle=

  27. [27]

    2020 , eprint=

    Scaling Laws for Neural Language Models , author=. 2020 , eprint=

    2020

  28. [28]

    Transactions on Machine Learning Research , year=

    Emergent Abilities of Large Language Models , author=. Transactions on Machine Learning Research , year=

  29. [29]

    International Conference on Learning Representations (ICLR) , year=

    Deep Double Descent: Where Bigger Models and More Data Hurt , author=. International Conference on Learning Representations (ICLR) , year=

  30. [30]

    2022 , eprint=

    Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets , author=. 2022 , eprint=

    2022

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    Gradient Surgery for Multi-Task Learning , author=. Advances in Neural Information Processing Systems , volume=

  32. [32]

    IEEE Robotics and Automation Letters , volume=

    Bayesian Gaussian Mixture Model for Robotic Policy Imitation , author=. IEEE Robotics and Automation Letters , volume=. 2019 , doi=

    2019

  33. [33]

    Black, Kevin and Brown, Noah and others , journal=

  34. [34]

    Robotics: Science and Systems (RSS) , year=

    Octo: An Open-Source Generalist Robot Policy , author=. Robotics: Science and Systems (RSS) , year=

  35. [35]

    Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and others , booktitle=

  36. [36]

    , journal=

    Zheng, et al. , journal=

  37. [37]

    Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (

    Wang, Lirui and Reed, Scott and others , booktitle=. Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (

  38. [38]

    , journal=

    Lyu, et al. , journal=

  39. [39]

    Robotics: Science and Systems (RSS) , year=

    Bridge Data: Boosting Generalization of Robotic Skills with Cross-Domain Datasets , author=. Robotics: Science and Systems (RSS) , year=

  40. [40]

    NIPS Workshop on Inductive Transfer: 10 Years Later , year=

    To Transfer or Not To Transfer , author=. NIPS Workshop on Inductive Transfer: 10 Years Later , year=

  41. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Characterizing and Avoiding Negative Transfer , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  42. [42]

    Wan, Weikang and others , journal=

  43. [43]

    arXiv preprint , year=

    Continual Learning for Robotic Manipulation: A Survey , author=. arXiv preprint , year=

  44. [44]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling Laws for Neural Language Models , author=. arXiv preprint arXiv:2001.08361 , year=

    Pith/arXiv arXiv 2001

  45. [45]

    Transactions on Machine Learning Research (TMLR) , year=

    Emergent Abilities of Large Language Models , author=. Transactions on Machine Learning Research (TMLR) , year=