Pith. sign in

REVIEW 3 major objections 3 minor 53 references

LAG-Fusion lets vision and force diffusion policies run asynchronously and fuse delay-compensated guidance, beating synchronous fusion on contact-rich manipulation.

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 18:31 UTC pith:RMJWJOHY

load-bearing objection The rebasing derivation is a real contribution, but the paper never specifies how the vision noise prediction is refreshed between 5 Hz queries, which makes the 25 Hz composition claim unverified. the 3 major comments →

arxiv 2607.17257 v1 pith:RMJWJOHY submitted 2026-07-19 cs.RO cs.AI

Asynchronous Multimodal Diffusion Policy Composition via Latency-Aware Guidance Fusion

Zihao He , Hongjie Fang , Shirun Tang , Cewu Lu , Haoshu Fang This is my paper
classification cs.RO cs.AI
keywords asynchronous multimodal policiesdiffusion policy compositionlatency-aware guidancereference-frame rebasecontact-rich manipulationforce feedbackreactive control
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 argues that multimodal diffusion policies fail to exploit high-frequency sensory feedback because they synchronize all modalities to the slowest branch. It proposes composing modality-specific diffusion policies asynchronously, letting each policy denoise at its native rate and contribute guidance whenever it is ready. To make delayed guidance comparable, it derives a reference-frame rebasing rule for relative action representations and fuses the aligned guidance with latency-aware weights that down-weight older forecasts. On two contact-rich tasks (flip a box and polish a curve), this approach outperforms synchronous composition, a learned-router consensus, and specialized force-aware policies while keeping inference at 25 Hz. If correct, the result gives a modular way to add fast reactive modalities to slow visual policies without retraining a monolithic network.

Core claim

The central claim is that asynchronous, latency-aware guidance fusion outperforms synchronous multimodal composition in contact-rich manipulation. The paper shows that a vision-conditioned diffusion policy and a force-conditioned diffusion policy can be composed at inference time even when queried at different times, by rebasing delayed vision predictions into the current force reference frame using an affine transform whose linear part is orthogonal, so Gaussian noise statistics are preserved, and then weighting the two guidance signals along the action horizon so that fresher force guidance dominates later steps. With this mechanism, the composed policy runs at 25 Hz and achieves the best

What carries the argument

The load-bearing object is the diffusion reference-frame rebase rule. For an action represented relative to the current robot pose, a frame change is affine: a' = A a + b, with A orthogonal. Under this transform, clean actions rebase affinely, noisy samples rebase with the offset scaled by sqrt(alpha_k), and predicted noises rebase only by the orthogonal part A. This preserves the standard Gaussian diffusion distribution while making delayed guidance expressible in the current frame. The second mechanism is the latency-aware weighting schedule: the vision weight decreases linearly from 0.8 to 0.2 over the fused horizon, treating older guidance as less reliable.

Load-bearing premise

The select-and-rebase composition requires the fast policy's action horizon to lie entirely inside the slow policy's horizon; if a fast policy predicts a longer chunk than the slow policy, the fusion rule cannot be applied.

What would settle it

Run the same two tasks with the latency-aware weights replaced by constant 0.5/0.5 or by the reversed schedule (vision weight increasing along the horizon); if success rates do not drop, the reported gains are not attributable to the latency-aware schedule.

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

If this is right

  • Modality-specific policies trained independently can be composed at test time, so adding a fast force or tactile branch does not require retraining a monolithic multimodal policy.
  • Synchronous fusion is identified as a primary cause of poor contact reactivity; any system bottlenecking fast feedback to the slowest sensing rate would suffer the same loss.
  • The rebase rule generalizes to any pair of diffusion policies with different latencies and overlapping horizons, not just vision-plus-force.
  • The latency-aware weighting provides a general confidence proxy: older or farther-ahead guidance is down-weighted, fresher feedback up-weighted.
  • The framework is instantiated with a 5 Hz vision policy and a 25 Hz force policy, showing the same interface could carry audio, tactile, or proprioceptive streams.

Where Pith is reading between the lines

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

  • The framework's modularity implies a practical scaling path: as new sensors are added to a robot, one could attach a new diffusion policy and fuse its guidance without disturbing existing modules; this consequence is not tested in the paper.
  • The hand-set weights (0.8 to 0.2 linear) are a candidate for learning; a learned or adaptive schedule could further improve performance, and an ablation of the schedule would clarify how much of the gain comes from the weights versus from asynchronous composition itself.
  • If the containment assumption (force horizon inside vision horizon) is relaxed, composition would require partial rebasing or a different selection operator; the rebase algebra suggests a generalization the paper leaves open.
  • The frequency analysis implies that any synchronous system will systematically underuse fast modalities; a direct test on tactile or audio streams could confirm this is a general principle, not just a force-policy effect.

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 proposes LAG-Fusion, a framework for composing modality-specific diffusion policies that run at different sensing/inference rates. It derives an affine reference-frame rebasing rule for relative action trajectories, showing that clean actions, noisy samples, and predicted noises transform differently under a pose change. The composed policy maintains a full-horizon noisy action in the vision frame, rebases the segment overlapping the force horizon into the force frame, fuses the rebased vision and force noise predictions with time-index-dependent weights, and maps the result back to the full horizon. Experiments on two contact-rich tasks (Flip and Polish Curve) compare against synchronous fusion, Policy Consensus, RDP, and TA-VLA, and a frequency study examines a standalone force policy. The paper claims that asynchronous, latency-aware fusion improves reactivity and task performance while preserving modularity.

Significance. If the central claims hold, LAG-Fusion addresses a real gap: existing diffusion-policy composition methods assume synchronous, same-frequency guidance. The reference-frame rebasing derivation in Section 3.2 is clear, parameter-free, and internally consistent; using the 6D rotation representation keeps the linear part orthogonal, which correctly preserves the Gaussian noise structure. The real-robot evaluation controls initial object configurations and includes a frequency study with unseen objects, which is a strength. However, the empirical and algorithmic support in the current manuscript is incomplete: the asynchronous loop is underspecified, the latency-weight schedule is unablated, and the main table has only 20 trials per condition. If these issues are resolved, the framework could be a useful modular alternative to synchronous multimodal diffusion policies.

major comments (3)
  1. [Section 3.3, Eqs. (8)–(10), Appendix A.3] The derivation assumes that ε^V_k and ε^F_k are evaluated at the same diffusion timestep k, but the deployment described in Table 1 and Appendix A.3 runs the composed policy at 25 Hz while the vision policy is queried at 5 Hz. The paper never specifies how the stored vision prediction ε^V_k is refreshed as k advances between vision queries. If the old ε^V is reused in Eq. (10), the composition is invalid because π^V depends on k and on the noisy action a_k. If ε^V is recomputed every 25 Hz step, then the vision branch is not operating at 5 Hz, undercutting the modularity/reactivity claim. Please provide the exact update rule for the DDIM trajectory at force-only steps and state how ε^V_k is maintained across asynchronous queries; this is load-bearing for the claimed 25 Hz asynchronous composition.
  2. [Appendix A.3, Eq. (11)] The latency-aware schedule w^V_i = 0.8 -> 0.2 is hand-selected; no ablation or sensitivity analysis is reported for w^V_start, w^V_end, or the linear shape. Since this schedule is the only 'latency-aware' component, the experiments do not establish that latency-aware weighting (rather than simply running the force branch asynchronously) is responsible for the gains over synchronous fusion. Add ablations with constant weights, reversed schedule, and variations of the endpoints.
  3. [Table 1 and Appendix B.1] The main comparative results are based on 20 trials per task. With n=20, the 60% vs 40% Flip success gap between LAG-Fusion and Policy Consensus is well within sampling error (roughly ±21 percentage points for a 95% binomial confidence interval), and no significance tests or repeated-seed variances are reported. This undermines the central empirical claim in Section 4.1. Report confidence intervals or additional trials, and describe how partial successes are counted in the percentages.
minor comments (3)
  1. [Section 3.3] The containment assumption 'the force horizon is contained within the vision horizon' is stated but never verified or discussed for the chosen horizons. Please state the vision action-chunk length in the experiments and discuss what happens if the fast policy's chunk extends beyond the slow policy's horizon.
  2. [Appendix A.2 and Table 2] The force policy predicts a 1 s action chunk at 50 Hz, while inference is 25 Hz. Clarify the relationship between action chunk timestep spacing and the control frequency, and how overlapping chunks are executed.
  3. [Figure 3 and Eq. (10)] The notation ε^{fuse}_k is used both before and after mapping back to T1; consider distinguishing the T2-frame fused noise from the final T1-frame noise.

Circularity Check

0 steps flagged

No significant circularity: the rebasing rule is a parameter-free algebraic derivation, and the hand-set latency weights are a hyperparameter, not a fitted prediction.

full rationale

LAG-Fusion's central claimed derivation is the reference-frame rebasing rule in Section 3.2. This is a direct algebraic consequence of the diffusion noisy-action definition (Eq. 1), the standard DDIM clean-action estimator (Eq. 5), and the affine relative-action frame transform (Eqs. 3-4). No fitted parameter is hidden in the rebase; the rule is parameter-free and would hold for any diffusion policy using relative actions and the DDIM estimator. The asynchronous composition in Section 3.3 uses this rule to align delayed guidance, and Eq. (10) is the standard weighted composition of two noise predictions after rebasing, not a statement equivalent to its output. The latency-aware weights in Eq. (11) are hand-selected hyperparameters (wV_start=0.8, wV_end=0.2, linear decay) rather than parameters fitted to the evaluation tasks, so the empirical claim is not forced by construction. The paper does contain self-citations (e.g., RISE [50], Force Policy [11]) and the vision backbone is drawn from the authors' prior work, but none of these citations functions as a load-bearing uniqueness theorem or ansatz; the compositional derivation and the external real-world task evaluations carry the argument. The reader-raised concern about diffusion-timestep alignment across asynchronous 5 Hz/25 Hz queries is a potential soundness or implementation gap, not a circularity: the paper does not reduce an output to an input, it simply leaves the control-loop update rule for k underspecified. Likewise, the stated assumption that the force horizon is contained within the vision horizon is an explicit limitation, not a circular premise. No specific equation or fitted parameter can be exhibited as equivalent to the conclusion, so the appropriate circularity verdict is no significant circularity (score 0).

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The central derivation depends on standard diffusion math plus a domain assumption about relative action frames and a segmentation assumption. The main hand-tuned inputs are the latency-aware weights, which are not ablated.

free parameters (3)
  • wV_start (vision weight at horizon start) = 0.8
    Hand-chosen in Appendix A.3, Eq. 11, with no ablation or optimization. Directly controls the fusion balance between vision and force guidance.
  • wV_end (vision weight at horizon end) = 0.2
    Hand-chosen complement such that wF = 1 - wV. No ablation reported; directly affects the composed noise and the reported task success.
  • Force policy architecture hyperparameters (MLP dims 600->256->16, U-Net channels [32,64], etc.)
    Standard network hyperparameters, not fitted to task success, but they determine the force policy's behavior and are not swept.
axioms (4)
  • standard math DDIM/DDPM noise schedule: a_k = sqrt(alpha_bar_k) a_0 + sqrt(1 - alpha_bar_k) epsilon (Eq. 1).
    Standard diffusion policy formulation, used as the starting point for the rebasing derivation.
  • domain assumption Relative action chunks are defined with respect to the robot pose at policy query time; the frame transform T_2<-1 = T_2^{-1} T_1 with orthogonal linear part A is valid for translation, 6D rotation columns, and gripper.
    Introduced in Section 3.2. It is the foundation of the rebase rule and is plausible for the chosen action representation, but is an assumption about how the robot and action space behave.
  • domain assumption The force policy horizon is temporally contained within the vision horizon.
    Explicitly assumed in Section 3.3 and central to the select-and-rebase procedure. Not justified beyond the specific instantiation (1s force chunk vs 1s+ vision chunk).
  • ad hoc to paper The composed noise epsilon^{fuse} replaces only the selected segment; the remaining full-horizon noise from the vision policy stays unchanged.
    Section 3.3 last paragraph. This is a design choice for composing guidance, not forced by the math. It could distort the diffusion trajectory at segment boundaries.

pith-pipeline@v1.3.0-alltime-deepseek · 13221 in / 7295 out tokens · 49430 ms · 2026-08-01T18:31:23.791881+00:00 · methodology

0 comments
read the original abstract

Diffusion policies have shown strong potential for robotic imitation learning, and recent extensions incorporate additional modalities to improve manipulation performance. However, these modalities often differ not only in information content but also in sensing rates and inference latencies. Existing multimodal diffusion policies typically rely on synchronous fusion or manually designed multi-frequency architectures, which either slow down high-frequency feedback or limit extensibility to new modality combinations. We propose LAG-Fusion, a latency-aware guidance fusion framework for asynchronous multimodal diffusion policy composition. LAG-Fusion allows modality-specific policies to operate at their native inference rates and contribute denoising guidance whenever available. To make asynchronous composition consistent, we derive a reference-frame rebasing rule for diffusion variables under relative action representations, enabling delayed guidance to be aligned before fusion. We instantiate LAG-Fusion in contact-rich manipulation by composing a low-frequency vision policy with a high-frequency force policy. Experiments under heterogeneous modality latencies show that LAG-Fusion improves policy responsiveness and task performance over synchronous fusion and specially designed force-aware baselines.

Figures

Figures reproduced from arXiv: 2607.17257 by Cewu Lu, Haoshu Fang, Hongjie Fang, Shirun Tang, Zihao He.

Figure 1
Figure 1. Figure 1: Overview of Multimodal Fusion Policy Paradigms. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the LAG-Fusion Framework. LAG-Fusion composes modality-specific diffusion policies at their native inference frequencies. (Left) A low-frequency vision policy provides long-horizon scene guidance, while a high-frequency force/torque policy provides local contact guidance. (Right) In the diffusion composition head, the overlapping noisy action segment is rebased into the current force frame for … view at source ↗
Figure 3
Figure 3. Figure 3: Latency-Aware Guidance Fusion. A low￾frequency vision policy provides full-horizon guidance in frame T1, while a high-frequency force policy is queried later in frame T2. LAG-Fusion selects the over￾lapping action segment, rebases it to T2 for force de￾noising, fuses the aligned vision and force guidance in T2, and maps the fused guidance back to T1 to update the full diffusion trajectory. We now apply the… view at source ↗
Figure 4
Figure 4. Figure 4: Contact-Rich Manipulation Tasks. We evaluate LAG-Fusion on two contact-rich tasks that require both visual perception and fine-grained force/torque feedback. Both tasks require effective coordination and adaptation across visual and force/torque modalities. The vision policy predicts full-horizon guidance ε V k = π V (ak ,k;c V ) in frame T1, where c V is the vision condition at time t1. To fuse it with th… view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation Objects for the Low-dimensional [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of Inference Frequency. (a) On Flip, the standalone low-dimensional force policy performs better when it is allowed to infer at higher frequencies, showing that the lightweight high-frequency modality benefits from its native high-frequency execution. (b) In the full LAG-Fusion framework, preserving high￾frequency inference also improves composed policy performance on both tasks. bottlenecking high-… view at source ↗
Figure 7
Figure 7. Figure 7: Overview of Demonstrations for each Tasks. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Evaluation Initial Configurations. For each task, we evaluate each policy on 20 different initial configurations. The object is randomly placed at the beginning of each trial, and these evaluation configurations are different from those used in the training demonstrations. action-chunking scheme-predicting and then executing a one-second action chunk per query-rather than from a limitation of its inference… view at source ↗
Figure 9
Figure 9. Figure 9: Evaluation Setup for the Standalone Low-dimensional Policy. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Detailed Standalone Force Policy Results under Different Input and Inference Frequencies. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] 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

53 extracted references · 6 linked inside Pith

  1. [1]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems, 2023

    2023

  2. [2]

    Janner, Y

    M. Janner, Y . Du, J. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis. InInternational Conference on Machine Learning, pages 9902–9915. PMLR, 2022

    2022

  3. [3]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    2020

  4. [4]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. InThe International Conference on Learning Representations, 2021

    2021

  5. [5]

    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. InRobotics: Science and Systems, 2025

    2025

  6. [6]

    H. Ha, P. Florence, and S. Song. Scaling up and distilling down: Language-guided robot skill acquisition. InConference on Robot Learning, pages 3766–3777. PMLR, 2023

    2023

  7. [7]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.Proceedings of Robotics: Science and Systems (RSS), 2024

    2024

  8. [8]

    Z. Liu, C. Chi, E. Cousineau, N. Kuppuswamy, B. Burchfiel, and S. Song. Maniwav: Learning robot manipulation from in-the-wild audio-visual data. InConference on Robot Learning, 2024

    2024

  9. [9]

    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

  10. [10]

    H. Chen, J. Xu, H. Chen, K. Hong, B. Huang, C. Liu, J. Mao, Y . Li, Y . Du, and K. Driggs- Campbell. Multi-modal manipulation via multi-modal policy consensus. In2026 IEEE Inter- national Conference on Robotics and Automation (ICRA), 2026. to appear, arXiv:2509.23468

    Pith/arXiv arXiv 2026

  11. [11]

    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

  12. [12]

    Y . Du, C. Durkan, R. Strudel, J. B. Tenenbaum, S. Dieleman, R. Fergus, J. Sohl-Dickstein, A. Doucet, and W. S. Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. InInternational conference on machine learning, pages 8489–8510. PMLR, 2023

    2023

  13. [13]

    Y . Du, S. Li, and I. Mordatch. Compositional visual generation and inference with energy based models. InAdvances in Neural Information Processing Systems, 2020

    2020

  14. [14]

    N. Liu, S. Li, Y . Du, A. Torralba, and J. B. Tenenbaum. Compositional visual generation with composable diffusion models. InComputer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, Lecture Notes in Computer Science, pages 423–439, 2022

    2022

  15. [15]

    L. Wang, J. Zhao, Y . Du, E. H. Adelson, and R. Tedrake. Policy composition from and for heterogeneous robot learning. InRobotics: Science and Systems, 2024. 9

    2024

  16. [16]

    J. Cao, Y . Huang, H. Guo, R. Zhang, M. Nan, W. Mai, J. Wang, H. Cheng, J. Sun, G. Han, W. Zhao, Q. Zhang, Y . Guo, Q. Zheng, C. Song, X. Li, P. Luo, and A. F. Luo. Compose your policies! improving diffusion-based or flow-based robot policies via test-time distribution-level composition. InInternational Conference on Learning Representations, 2026

    2026

  17. [17]

    C. Liu, H. Chen, S. H. Høeg, S. Yao, Y . Li, K. Hauser, and Y . Du. Flexible multitask learning with factorized diffusion policy.IEEE Robotics and Automation Letters, 11(4):4697–4704, 2026

    2026

  18. [18]

    D. Geng, I. Park, and A. Owens. Visual anagrams: Generating multi-view optical illusions with diffusion models. InConference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  19. [19]

    D. Geng, I. Park, and A. Owens. Factorized diffusion: Perceptual illusions by noise decompo- sition. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  20. [20]

    Z. Chen, D. Geng, and A. Owens. Images that sound: Composing images and sounds on a single canvas.Neural Information Processing Systems (NeurIPS), 2024

    2024

  21. [21]

    A. Liu, M. Sap, X. Lu, S. Swayamdipta, C. Bhagavatula, N. A. Smith, and Y . Choi. Dexperts: Decoding-time controlled text generation with experts and anti-experts. InAnnual Meeting of the Association for Computational Linguistics, 2021

    2021

  22. [22]

    Yang and D

    K. Yang and D. Klein. Fudge: Controlled text generation with future discriminators. InCon- ference of the North American Chapter of the Association for Computational Linguistics, 2021

    2021

  23. [23]

    Krause, A

    B. Krause, A. D. Gotmare, B. McCann, N. S. Keskar, S. Joty, R. Socher, and N. F. Rajani. Gedi: Generative discriminator guided sequence generation. InFindings of the Association for Computational Linguistics: EMNLP, 2021

    2021

  24. [24]

    X. Lu, P. West, R. Zellers, R. Le Bras, C. Bhagavatula, and Y . Choi. Neurologic decoding: (un)supervised neural text generation with predicate logic constraints. InConference of the North American Chapter of the Association for Computational Linguistics, 2021

    2021

  25. [25]

    U. A. Mishra, S. Xue, Y . Chen, and D. Xu. Generative skill chaining: Long-horizon skill planning with diffusion models. InConference on Robot Learning, 2023

    2023

  26. [26]

    U. A. Mishra, D. He, Y . Chen, and D. Xu. Compositional diffusion with guided search for long-horizon planning. InInternational Conference on Learning Representations, 2026

    2026

  27. [27]

    Y . Du, M. Yang, P. Florence, F. Xia, A. Wahid, B. Ichter, P. Sermanet, T. Yu, P. Abbeel, J. B. Tenenbaum, C. Lynch, and J. Tompson. Energy-based models are zero-shot planners for compositional scene rearrangement. InRobotics: Science and Systems, 2023

    2023

  28. [28]

    N. Hogan. Impedance control: An approach to manipulation.Journal of Dynamic Systems, Measurement, and Control, 107:1–24, 1985

    1985

  29. [29]

    Buchli, F

    J. Buchli, F. Stulp, E. Theodorou, and S. Schaal. Learning variable impedance control.Inter- national Journal of Robotics Research, 30(7):820–833, 2011

    2011

  30. [30]

    H. Seraji. Adaptive admittance control: An approach to explicit force control in compliant motion. InIEEE International Conference on Robotics and Automation, pages 2705–2712. IEEE, 1994

    1994

  31. [31]

    Kronander and A

    K. Kronander and A. Billard. Stability considerations for variable impedance control.IEEE Transactions on Robotics, 32(5):1298–1305, 2016

    2016

  32. [32]

    M. H. Raibert and J. J. Craig. Hybrid position/force control of manipulators.Journal of dynamic systems, measurement, and control, 103(2):126–133, 1981. 10

    1981

  33. [33]

    O. Khatib. A unified approach for motion and force control of robot manipulators: The opera- tional space formulation.IEEE Journal on Robotics and Automation, 3(1):43–53, 2003

    2003

  34. [34]

    H. Li, Y . Zhang, J. Zhu, S. Wang, M. A. Lee, H. Xu, E. H. Adelson, L. Fei-Fei, R. Gao, and J. Wu. See, hear, and feel: Smart sensory fusion for robotic manipulation. InConference on Robot Learning, pages 1368–1378, 2022

    2022

  35. [35]

    Mejia, V

    J. Mejia, V . Dean, T. Hellebrekers, and A. Gupta. Hearing touch: Audio-visual pretraining for contact-rich manipulation.arXiv preprint arXiv:2405.08576, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    R. Feng, D. Hu, W. Ma, and X. Li. Play to the score: Stage-guided dynamic multi-sensory fusion for robotic manipulation. InConference on Robot Learning, 2024

    2024

  37. [37]

    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. In Advances in Neural Information Processing Systems, 2025

    2025

  38. [38]

    Zhang, H

    Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. Elucidating the design space of torque-aware vision-language-action models. InConference on Robot Learning, volume 305, pages 4019–4037. PMLR, 2025

    2025

  39. [39]

    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. InIEEE International Conference on Robotics and Automation, pages 4829–4836, 2025

    2025

  40. [40]

    X. Xu, Y . Hou, Z. Liu, and S. Song. Compliant residual DAgger: Improving real-world contact- rich manipulation with human corrections. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  41. [41]

    H.-S. Fang, B. Romero, Y . Xie, A. Hu, B.-R. Huang, J. Alvarez, M. Kim, G. Margolis, K. An- barasu, M. Tomizuka, E. Adelson, and P. Agrawal. Dexop: A device for robotic transfer of dexterous human manipulation.arXiv preprint arXiv:2509.04441, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    K. Yu, Y . Han, Q. Wang, V . Saxena, D. Xu, and Y . Zhao. Mimictouch: Leveraging multi-modal human tactile demonstrations for contact-rich manipulation. InConference on Robot Learning, 2024

    2024

  43. [43]

    Zhang, J

    Z. Zhang, J. Ma, X. Yang, X. Wen, Y . Zhang, B. Li, Y . Qin, J. Liu, C. Zhao, L. Kang, et al. Touchguide: Inference-time steering of visuomotor policies via touch guidance.arXiv preprint arXiv:2601.20239, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    J. Yin, H. Qi, Y . Wi, S. Kundu, M. Lambeta, W. Yang, C. Wang, T. Wu, J. Malik, and T. Helle- brekers. Osmo: Open-source tactile glove for human-to-robot skill transfer.arXiv:2512.08920, 2025

    arXiv 2025

  45. [45]

    S. Xia, H. Fang, H.-S. Fang, and C. Lu. Cage: Causal attention enables data-efficient general- izable robotic manipulation. InIEEE International Conference on Robotics and Automation. IEEE, 2025

    2025

  46. [46]

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. InRobotics: Science and Systems, 2024

    2024

  47. [47]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InRobotics: Science and Systems, 2023

    2023

  48. [48]

    Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li. On the continuity of rotation representations in neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745–5753, 2019. 11

    2019

  49. [49]

    F. Ltd. Flexiv teleoperation development kit (tdk). Accessed January 2026

    2026

  50. [50]

    C. Wang, H. Fang, H.-S. Fang, and C. Lu. Rise: 3d perception makes real-world robot imitation simple and effective.arXiv preprint arXiv:2404.12281, 2024

    Pith/arXiv arXiv 2024

  51. [51]

    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. InICRA, pages 1105–1112. IEEE, 2025

    2025

  52. [52]

    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. InRobotics: Science and Systems, 2025

    2025

  53. [53]

    Black, N

    K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y . Galliker, et al.π0.5: a vision-language-action model with open-world general- ization. InConference on Robot Learning, volume 305, pages 17–40. PMLR, 2025. 12 Supplementary Material A Method Implementation Details A.1 Visual Policy Details Our visual...

    2025