Pith. sign in

REVIEW 3 major objections 3 minor 29 references

LIFT shows that a pretrained vision-language-action policy can gain contact reactivity by injecting recent 6D force through a zero-initialized reactive action expert, learning faster and reaching higher success than vision-only post-trainin

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

2026-08-02 02:40 UTC pith:DGK2PXLE

load-bearing objection Clever initialization-equivalent architecture for adding force to a VLA, but the empirics don't yet isolate force from the reactive branch. the 3 major comments →

arxiv 2607.14236 v1 pith:DGK2PXLE submitted 2026-07-15 cs.RO cs.AIcs.LG

Never Too Late for Force: Accelerating VLA Post-Training with Reactive Force Injection

Yi Wang , Wendi Chen , Zimo Wen , Han Xue , Xueqi Li , Wenye Yu , Zhijie Chen , Hao Yang
show 3 more authors
Jun Lv Chuan Wen Cewu Lu
This is my paper
classification cs.RO cs.AIcs.LG
keywords vision-language-action modelsforce feedbackcontact-rich manipulationonline post-trainingDAggerreactive controlrobot manipulationimitation learning
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 sets out to show that a pretrained vision-language-action (VLA) model, trained on vision and language alone, can be made reactive to physical contact after the fact — without losing what it already knows. It proposes LIFT, which grafts a second 'reactive action expert' beside the original action decoder, feeds recent 6D end-effector force through a causal memory, and injects that force through zero-initialized cross attention so the model initially produces exactly the same actions as before. Coupled with an online DAgger loop that collects human corrections on states the robot actually visits, LIFT reports faster learning and higher final success than vision-only post-training on three contact-rich real-robot tasks — towel folding, book insertion, and Hanoi ring placement — while preserving the base model's generalization under object, lighting, and tablecloth changes. The paper's underlying claim is that force can be added late, as a reactive residual, rather than baked in during large-scale pretraining.

Core claim

The central claim is that force-aware post-training does not need to start from a force-trained model: a reactive action expert copied from a pretrained VLA's action head, combined with a causal force memory and zero-initialized cross attention, can refresh individual actions within an action chunk as new 6D wrench measurements arrive, without disturbing the pretrained prior at initialization. On the three evaluated tasks, LIFT reaches peak scores of 0.825, 0.6, and 0.6 against 0.725, 0.4, and 0.3 for vision-only online DAgger, while using fewer or comparable samples. The author's way of putting it is that force injection accelerates post-training, and that reactive force memory — not single

What carries the argument

The load-bearing object is the reactive action expert: a second action-denoising stream that shares the pretrained weights but uses shifted causal attention, so each action token can see earlier reactive tokens and later base-action tokens without seeing past base tokens. A causal GRU encodes the recent 6D wrench into a force memory, and zero-initialized cross attention lets that memory add a residual update that is exactly zero before training. The latency-aligned causal mask (offset L in the paper) ensures each action can only use force that would actually be available when inference finishes. Together these pieces make the augmented policy output-equivalent to the original VLA at step zer

Load-bearing premise

The hand-set latency-alignment offset L=3 is assumed to exactly match the robot's inference delay, so the force memory is neither stale nor future when an action is generated; the paper reports no sensitivity analysis over L, and a wrong offset would break the very reactivity LIFT is designed for.

What would settle it

Vary L over 0, 1, 2, 3, 4, 5 on the same three tasks while keeping everything else fixed; if success does not degrade sharply at the wrong offsets, the latency-alignment assumption is not load-bearing. A direct check is to measure the actual network inference delay and confirm that L=3 aligns the force measurement time with the action execution time.

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

If this is right

  • Force can be injected into a VLA late in post-training; a pretrained visual prior does not have to be discarded to gain contact reactivity.
  • Reactive force memory outperforms single-frame force inputs, because contact-phase decisions like 'still inserting vs. already bottomed out' require a short force history.
  • Online DAgger is necessary: the same reactive model trained offline only drops to zero success on book insertion, showing that on-policy contact data cannot be replaced by offline demonstrations.
  • The augmented policy retains the base VLA's generalization: performance holds under object, tablecloth, and lighting changes.
  • Because the force path starts at zero output, the method adds force capability without requiring an additional force-pretraining stage.

Where Pith is reading between the lines

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

  • If the reported gains hold up, the same residual-injection pattern could be reused for other high-frequency, low-dimensional signals that are expensive to precollect — joint torques, tactile arrays, or contact microphones — turning LIFT into a general 'add a sensor late' recipe.
  • The heavy reliance on the hand-set latency offset L=3 suggests that deployment on a different arm or policy speed would require either a principled way to set L from measured inference latency or a mechanism that self-corrects misalignment; this is a testable engineering question the paper leaves open.
  • Because the method trains on a 1:1 mix of offline vision data and online force corrections, it implicitly argues that platform-specific force data can be scarce if paired with abundant visual demonstrations; an interesting test would be sweeping the online-data fraction to find the minimum correction count per task.
  • One unstated consequence is that force sensing can remain a post-training-only hardware requirement: the same pretrained checkpoint can be deployed on force-equipped and non-force arms, with the latter falling back to the base action stream.

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

3 major / 3 minor

Summary. The paper introduces LIFT, a post-training framework that adds reactive 6D force feedback to a pretrained vision-language-action (VLA) policy. A reactive action expert is grafted beside the original action expert, initialized by copying the pretrained action weights, with force injected through a causal GRU-based force memory and zero-initialized cross attention. Shifted causal attention is designed so that the reactive stream is output-equivalent to the original policy at initialization. Post-training uses a mixture of vision-only offline data and force-enabled online DAgger corrections. The empirical evaluation covers three real-robot tasks (towel folding, book insertion, Hanoi ring placement), comparing LIFT against vision-only online DAgger, a single-frame-force variant, an offline-only variant, and an offline handheld baseline. The paper reports faster learning and higher final success rates for LIFT on all three tasks, with ablations showing the importance of reactive force memory and online DAgger.

Significance. If the central empirical claim held, LIFT would be a useful contribution: it gives a concrete, architecture-level recipe for adding contact reactivity to a general VLA without destroying the pretrained prior. The initialization-equivalence construction (shifted causal attention plus zero-initialized cross attention, Section 3.2) is elegant and, as described, plausible; this is a genuine conceptual strength. The online DAgger post-training loop with mixed offline/online sampling is practical and the evaluation on three real-robot contact-rich tasks is valuable. However, the paper's headline conclusion—that force accelerates and improves VLA post-training—is not yet established because the main comparison confounds force injection with a major architectural change, and the statistical support is weak (10 rollouts/checkpoint, no error bars). The contributions are therefore significant but provisional.

major comments (3)
  1. [Section 4.1/4.2] The primary comparison does not isolate the effect of force. 'π0.5 w/ Online DAgger' is described as 'the same online DAgger loop without force input,' but it is the original π0.5 policy; it lacks the reactive action expert (O1.1), the shifted causal attention (O2.1), the cached vision-language prefix (O1.3), and the zero-initialized cross attention. Thus the headline 'LIFT learns faster and reaches higher performance than vision-only post-training' compares force+reactive architecture against a non-reactive open-loop architecture. The ablation 'LIFT w/o Reactive Force Injection' retains the reactive expert but uses single-frame force, so it does not remove force. On towel folding (Section 4.2) this ablation reaches 0.95 versus LIFT's 0.825, indicating that non-force reactive components alone can be highly effective. Without a no-force reactive control—i.e., a variant that keeps the reac
  2. [Appendix G / Figure 5] All performance numbers are means over 10 autonomous rollouts per checkpoint, with no variance, confidence intervals, or significance tests. With n=10, the reported differences such as 0.6 vs 0.4 (book insertion) are within ordinary binomial sampling noise; e.g., 6/10 vs 4/10 have overlapping 95% confidence intervals. The same concern applies to the Hanoi numbers (0.6 vs 0.3) and towel folding (0.825 vs 0.725). Please report per-rollout scores, include error bars or CIs on the curves, and provide a statistical test (e.g., bootstrap or exact binomial) for the key comparisons. Without this, the central claim of higher final performance is not supported at the stated sample size.
  3. [Eq. (3) / Table 2] The latency-alignment offset L=3 is load-bearing for the reactivity claim: it determines which force measurements are available to each reactive action token. The paper states only that it is 'chosen to match the network inference delay' but gives no measurement procedure and no sensitivity analysis. If L is mis-set, the force memory is either stale or uses force that is not yet available, defeating the purpose of the causal mask. Please report how L was estimated (e.g., end-to-end inference latency profiling on the Flexiv Rizon 4S), and vary L (e.g., 1, 2, 3, 4) on at least one task to show that performance is not highly sensitive to this hand-set value.
minor comments (3)
  1. [Section 4.4 / Figure 6] The generalization claim ('no clear drop') is qualitative. Please provide the numerical success rates for each shifted condition and, ideally, confidence intervals, so readers can verify that the differences are not just within noise.
  2. [Section 3.2 / Appendix B] The output-equivalence argument is stated informally. Since it is central to the prior-preservation claim, consider stating it as a short formal proposition (with the attention masking pattern of Figure 3) and proving it in the appendix. This would make the argument easier to audit.
  3. [Algorithm 1 / O1.2] The paper consistently refers to 'causal force memory,' but Algorithm 1 clears the force history at the start of each chunk (line 6) and Eq. (2) encodes only F_{t:t+H}. Clarify whether memory spans only the current chunk or a longer history; the term 'memory' suggests longer persistence.

Circularity Check

0 steps flagged

No significant circularity: LIFT's headline claims are empirically measured; its initialization guarantee is a construction, not a fitted prediction.

full rationale

The paper does not offer a derivation that could reduce to its own inputs. The headline result—LIFT learns faster and reaches higher performance than vision-only DAgger—is an empirical comparison on real-robot rollouts (§4.2, Figure 5); success rates are measured in the world, not manufactured by normalization or by a fitted equation. The load-bearing initialization claim (Section 3.2) is an architectural construction: copying base action-expert weights into the reactive expert and zero-initializing the cross-attention output projection. Even if the output-equivalence argument were imperfect, it is not a circular use of the experimental outcome; it is an initial-condition design. Equation (1) is a standard flow-matching objective over both streams, and Eq. (3) introduces a latency-alignment offset L=3 with no fitted relation to the reported success rates. Many references are from the same group (e.g., [7]–[13], [16], [21]), but they are used as related work or as architectural/system inspiration (e.g., SOP's adaptive sampler [15]), not as the source of the measured gains. The stated limitations (human corrections, single-arm experiments) do not assert any circular step. A possible confound—comparing LIFT to a vision-only baseline that lacks the reactive action branch—is an experimental design issue, not a circularity under the requested definition. Therefore no circular step can be quoted, and the score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central empirical result (LIFT beats vision-only DAgger) depends on the architecture working as described, which in turn rests on the init-equivalence construction, the latency offset, force data quality, and DAgger assumptions. The paper provides detailed specifications but no formal verification or sensitivity analysis for these premises.

free parameters (3)
  • latency-alignment offset L = 3
    Hand-set to match the network inference delay (Eq. 3, Table 2). No sensitivity analysis is provided; the reactive property of LIFT depends on this offset being correct for the deployment platform.
  • DAgger adaptive sampling exponent alpha = 1.5
    Chosen for the online/offline mixture in Appendix H; not tuned or ablated, and affects the balance between prior preservation and online correction.
  • force encoder widths = d_h=512, d_m=1024
    Architecture capacity chosen by hand (Table 1); central claim is not obviously sensitive to these, but they are not derived from data.
axioms (5)
  • standard math Softmax attention with a zero-initialized output projection adds exactly zero residual at initialization.
    Used in O2.2 (Section 3.2) to guarantee the force path does not change action outputs at step 0.
  • ad hoc to paper Under the copied weights and shifted causal attention, the reactive token r_i is context-equivalent to base token a_i at initialization, making the two streams output-equivalent.
    Asserted in Section 3.2 (O2.1, Figure 3); argued informally but not formally proven or empirically verified in the paper.
  • domain assumption The pretrained pi0.5 VLA provides a useful general manipulation prior worth preserving during post-training.
    The whole premise of 'late force injection' (Section 1, Problem Formulation) assumes the base policy's knowledge transfers to the target tasks and should be kept.
  • domain assumption 6D wrench sampled at 10 Hz is sufficiently informative and correctly synchronized with the action stream for contact reactivity.
    Hardware setup (Appendix D) assumes the downsampled force signal is accurate and time-aligned; errors here would break the force memory.
  • domain assumption Online DAgger corrections collected on learner-distribution states improve the policy and mitigate covariate shift.
    The training recipe (O3.3, Appendix H) relies on the standard DAgger premise that on-policy corrective data reduce distribution shift.

pith-pipeline@v1.3.0-alltime-deepseek · 14836 in / 15249 out tokens · 157009 ms · 2026-08-02T02:40:54.975407+00:00 · methodology

0 comments
read the original abstract

Pretrained vision-language-action (VLA) policies provide strong language-conditioned manipulation knowledge, but they remain largely vision-driven and can struggle once manipulation enters contact states where the scene is occluded, depth is ambiguous, or small force errors push execution off the offline demonstration distribution. We present LIFT (Late Reactive Injection of Force for VLA Post-Training), a force-aware post-training framework that adds contact reactivity to a pretrained VLA policy while preserving its general manipulation knowledge. LIFT grafts a reactive action expert beside the original action expert, initializes it from pretrained action weights, and injects recent 6D end-effector force through causal force memory and zero-initialized cross attention, enabling actions to be refreshed during execution. To address the policy-dependent distribution shift of contact feedback, LIFT further couples reactive force injection with an online DAgger loop that trains on a mixture of offline task-alignment data and human-corrected online rollouts. Across towel folding, book insertion, and Hanoi ring placement, LIFT learns faster and reaches higher performance than vision-only post-training, while ablations show that reactive force memory and online corrective data are both important for robust contact-rich manipulation. Our code and data will be publicly available.

Figures

Figures reproduced from arXiv: 2607.14236 by Cewu Lu, Chuan Wen, Han Xue, Hao Yang, Jun Lv, Wendi Chen, Wenye Yu, Xueqi Li, Yi Wang, Zhijie Chen, Zimo Wen.

Figure 1
Figure 1. Figure 1: LIFT overview. LIFT adds reactive force control to a pretrained VLA by encoding recent force as causal force memory and injecting it through cross attention for within-chunk action updates. To preserve the pretrained prior, the reactive action expert is copied from the base action expert and the force-injected cross attention is zero-initialized, making the augmented policy initially output-equivalent to t… view at source ↗
Figure 2
Figure 2. Figure 2: LIFT architecture. LIFT grafts a reactive action expert beside pretrained π0.5 while keeping the original vision-language stack unchanged. Recent force is encoded as force memory and injected into the reactive branch through zero-initialized cross attention, enabling within-chunk action updates while preserving the pretrained prior at initialization. alignment dataset Dv = {(ℓ, I, a)}, from which a vision-… view at source ↗
Figure 3
Figure 3. Figure 3: Shifted causal attention with ini￾tialization equivalence. Each reactive token attends to the vision-language prefix, later base-action tokens, and the causal reactive prefix. Combined with the copied pretrained action expert, LIFT preserves initialization equivalence while enabling causal reactive updates. O2.1 Make reactive action expert output￾equivalent. Starting from pretrained π0.5, we instantiate a … view at source ↗
Figure 4
Figure 4. Figure 4: LIFT training pipeline. The two-stage system first aligns LIFT on vision-only demon￾strations and then closes an online post-training loop between the robot and cloud trainer. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance curves on three tasks as sample count increases. By leveraging reactive control and force memory, LIFT reaches higher peak performance with fewer samples. 4 Experimental Results 4.1 Experimental setup We evaluate on three real-robot manipulation tasks: towel folding, book insertion, and Hanoi ring placement. The stage-one vision-only data are collected with a handheld data-collection device, wh… view at source ↗
Figure 6
Figure 6. Figure 6: Generalization of the final LIFT checkpoint. The final LIFT checkpoint maintains similar performance under object, tablecloth, and lighting changes across the three tasks. Hanoi ring placement shows the same pattern in a precision placement setting. LIFT reaches 0.6 (# 1.7K) and maintains 0.6 at the final checkpoint (# 2.0K), whereas π0.5 w/ Online DAgger peaks at 0.3 (# 1.4K) and ends at 0.2 (# 1.6K). Her… view at source ↗
Figure 7
Figure 7. Figure 7: Representative success and failure cases for the online DAgger ablation. They show that force injection and reactive control lead to more robust behavior during online post-training. 4.5 Does online data matter for reactive force injection? (Q4) Offline-only comparisons are plotted as separate markers in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generalization task settings. The three panels show the towel folding, book insertion, and Hanoi ring placement setups used for shifted-condition evaluation. Towel folding. • Blue towel: replace the original towel with a blue towel of the same style. • Another tablecloth: replace the rough cartoon-style brown tablecloth with a slightly smoother yellow tablecloth with an oil-painting texture. • Purple light… 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

29 extracted references · 11 linked inside Pith

  1. [1]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023

    2023

  2. [2]

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.π 0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    G. A. Team. Gen-1: Scaling embodied foundation models to mastery.Generalist AI Blog,

  7. [7]

    Z. He, H. Fang, J. Chen, H.-S. Fang, and C. Lu. Foar: Force-aware reactive policy for contact- rich robotic manipulation.IEEE Robotics and Automation Letters, 2025

    2025

  8. [8]

    W. Liu, J. Wang, Y . Wang, W. Wang, and C. Lu. Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 1105–1112. IEEE, 2025

    2025

  9. [9]

    Y . Hou, Z. Liu, C. Chi, E. Cousineau, N. Kuppuswamy, S. Feng, B. Burchfiel, and S. Song. Adaptive compliance policy: Learning approximate compliance for diffusion guided control. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4829–

  10. [10]

    H. Xue, J. Ren, W. Chen, G. Zhang, Y . Fang, G. Gu, H. Xu, and C. Lu. Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation.arXiv preprint arXiv:2503.02881, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    W. Chen, H. Xue, Y . Wang, F. Zhou, J. Lv, Y . Jin, S. Tang, C. Wen, and C. Lu. Implicitrdp: An end-to-end visual-force diffusion policy with structural slow-fast learning.arXiv preprint arXiv:2512.10946, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    H. Yuan, W. Yi, Z. Zhang, W. Chen, Y . Mo, J. Yin, X. Li, X. Zeng, C. Wen, C. Lu, K. Driggs- Campbell, and I. Lourentzou. Vtam: Video-tactile-action models for complex physical inter- action beyond vlas.arXiv preprint arXiv:2603.23481, 2026

    arXiv 2026

  13. [13]

    C. Yuan, Z. Zhang, M. Zhou, W. Chen, Y . Wang, Z. Liu, D. Niu, S. Wang, H. Zhang, W. Zhang, Y . Hu, Y . Gong, W. Xing, C. Wen, C. Lu, K. Zhang, and Y . Gao. Ftp-1: A generalist foundation tactile policy across tactile sensors for contact-rich manipulation.arXiv preprint arXiv:2606.13102, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    P. J. Ball, L. Smith, I. Kostrikov, and S. Levine. Efficient online reinforcement learning with offline data. InInternational Conference on Machine Learning, pages 1577–1594. PMLR, 2023. 9

    2023

  15. [15]

    M. Pan, S. Feng, Q. Zhang, X. Li, J. Song, C. Qu, Y . Wang, C. Li, Z. Xiong, Z. Chen, et al. Sop: A scalable online post-training system for vision-language-action models.arXiv preprint arXiv:2601.03044, 2026

    arXiv 2026

  16. [16]

    J. Fang, W. Chen, H. Xue, F. Zhou, T. Le, Y . Wang, Y . Zhang, J. Lv, C. Wen, and C. Lu. Robopocket: Improve robot policies instantly with your phone.arXiv preprint arXiv:2603.05504, 2026

    arXiv 2026

  17. [17]

    Flexiv TDK: Teleoperation development kit for flexiv robots.https:// github.com/flexivrobotics/flexiv_tdk, 2026

    Flexiv Robotics. Flexiv TDK: Teleoperation development kit for flexiv robots.https:// github.com/flexivrobotics/flexiv_tdk, 2026. GitHub repository

    2026

  18. [18]

    J. J. Liu, Y . Li, K. Shaw, T. Tao, R. Salakhutdinov, and D. Pathak. Factr: Force-attending curriculum training for contact-rich policy learning.arXiv preprint arXiv:2502.17432, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    C. Chen, Z. Yu, H. Choi, M. Cutkosky, and J. Bohg. Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation.IEEE Robotics and Au- tomation Letters, 2025

    2025

  20. [20]

    Zhang, H

    Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. Ta-vla: Elu- cidating the design space of torque-aware vision-language-action models.arXiv preprint arXiv:2509.07962, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    H. Fang, S. Tang, M. Mei, H. Qin, Z. He, J. Chen, Y . Feng, C. Wang, W. Liu, Z. He, C. Lu, and S. Wang. Force policy: Learning hybrid force-position control policy under interaction frame for contact-rich manipulation.arXiv preprint arXiv:2602.22088, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai, et al. Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation.Ad- vances in Neural Information Processing Systems, 38:93409–93439, 2026

    2026

  23. [23]

    Y . Li, H. Jiang, J. Xia, H. Zhang, J. Du, Y . Zhou, J. Zeng, C. Hao, J. Ren, Q. Yu, et al. Forcevla2: Unleashing hybrid force-position control with force awareness for contact-rich ma- nipulation.arXiv preprint arXiv:2603.15169, 2026

    arXiv 2026

  24. [24]

    S. Ross, G. Gordon, and D. Bagnell. A reduction of imitation learning and structured predic- tion to no-regret online learning. InProceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627–635. JMLR Workshop and Conference Pro- ceedings, 2011

    2011

  25. [25]

    Kelly, C

    M. Kelly, C. Sidrane, K. Driggs-Campbell, and M. J. Kochenderfer. Hg-dagger: Interactive imitation learning with human experts. In2019 International Conference on Robotics and Automation (ICRA), pages 8077–8083. IEEE, 2019

    2019

  26. [26]

    H. Liu, S. Nasiriany, L. Zhang, Z. Bao, and Y . Zhu. Robot learning on the job: Human-in- the-loop autonomy and learning during deployment.The International Journal of Robotics Research, 44(10-11):1727–1742, 2025

    2025

  27. [27]

    H. Liu, Y . Zhang, V . Betala, E. Zhang, J. Liu, C. Ding, and Y . Zhu. Multi-task interactive robot fleet learning with visual world models.arXiv preprint arXiv:2410.22689, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    X. Xu, Y . Hou, Z. Liu, and S. Song. Compliant residual dagger: Improving real-world contact- rich manipulation with human corrections.Advances in Neural Information Processing Sys- tems, 38:139559–139581, 2026. 10 A Full Related Work Force-augmented policies.Contact-rich manipulation work increasingly models force and torque explicitly rather than relyin...

    2026

  29. [2026]

    https://generalistai.com/blog/apr-02-2026-GEN-1

    2026