Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

This paper proposes that a single end-to-end diffusion policy can fuse slow vision and fast force into one causal transformer, replacing the hierarchical slow-fast design, and that a virtual-target auxiliary objective stops modality collaps

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

2026-08-03 16:58 UTC pith:ECW4QGCZ

load-bearing objection Architecturally real step forward on slow-fast visual-force diffusion; the empirical claim over RDP is thinner than the headline suggests, especially on box flipping. the 4 major comments →

arxiv 2512.10946 v2 pith:ECW4QGCZ submitted 2025-12-11 cs.RO cs.AIcs.LG

ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast Learning

Wendi Chen , Han Xue , Yi Wang , Fangyuan Zhou , Jun Lv , Yang Jin , Shirun Tang , Chuan Wen
show 1 more author
Cewu Lu
This is my paper
classification cs.RO cs.AIcs.LG
keywords imitation learningforce controldiffusion policycontact-rich manipulationmulti-modal policycausal attentionvirtual targetslow-fast 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.

Contact-rich manipulation requires both slow, spatial vision and fast, local force feedback, but existing policies either ignore force or delegate it to a separate hierarchy. This paper argues that a single end-to-end diffusion policy can close the loop on force inside an action chunk: a transformer with causal attention interleaves low-frequency visual tokens and high-frequency force tokens, and a cached deterministic DDIM sampler lets the same network replan per control step using the latest force. To keep the network from relying on one modality alone, it predicts a 'virtual target' — a compliance-based, action-space image of the desired force — as an auxiliary objective. If the reported results hold, a unified network can outperform both vision-only and hierarchical baselines while simplifying training.

Core claim

The central claim is that the slow-fast split does not require two networks. By using a gated-recurrent encoder for force tokens and a causal attention mask so that each action token sees only past forces, the policy can generate an action chunk and then extend it one token at a time while reusing the same initial noise and slow observation cache, with DDIM stochasticity set to zero for consistency. The network is trained to denoise an augmented action vector that concatenates the action, the virtual target x_vt = x_real + K^-1 f_ext, and a stiffness magnitude, with adaptive stiffness to up-weight contact events. The paper reports 18/20 successes on both box flipping and switch toggling, aga

What carries the argument

The load-bearing mechanism is temporal causality inside a Transformer-based diffusion policy. A gated-recurrent encoder processes force tokens in order, and a causal attention mask stops each action token from seeing future forces. At inference, DDIM with eta=0 and cached slow tokens and initial noise makes consecutive denoising passes deterministic, so the executed action at each step is a function of the latest force. The auxiliary virtual-target prediction maps measured force into the action coordinate frame through a quasi-static compliance model with adaptive stiffness, giving a learning signal that shares the action space and up-weights large contact forces.

Load-bearing premise

The virtual-target training signal rests on a simplified quasi-static contact model with hand-chosen stiffness values; if that model does not match real contact forces, the claimed advantage over force prediction may not hold.

What would settle it

Run the same policy on a task with significant dynamic contact, such as a high-speed impact, and compare ImplicitRDP against a variant trained with raw force prediction; if the virtual-target advantage disappears, the quasi-static assumption is load-bearing. Simpler: fix stiffness to a constant and see whether the 18/20 result changes.

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

If this is right

  • Force-based closed-loop control can be achieved within action chunking without a separate fast policy, preserving smooth, non-Markovian action chunks.
  • Representing force in the action space via virtual targets may generalize better than raw force prediction as an auxiliary objective.
  • Temporal causality plus consistent denoising is a reusable recipe for injecting any high-frequency observation into a diffusion policy.
  • End-to-end trained visual-force policies can outperform hierarchical designs in contact-rich manipulation with a simpler pipeline.

Where Pith is reading between the lines

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

  • The cached-noise, deterministic-DDIM inference scheme should transfer to other diffusion policies with causal structure, letting tactile or audio streams be injected at their native rate without retraining the architecture.
  • The virtual-target objective is a per-step force-to-displacement mapping; if folded into vision-language-action models, it could serve as a physics-grounded alternative to future-torque prediction without extra task-specific losses.
  • A natural stress test would vary the stiffness thresholds (f_min, f_max, k_min, k_max) and measure success-rate sensitivity; the paper does not report these values.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. ImplicitRDP proposes a single end-to-end visual-force diffusion policy for contact-rich manipulation. It replaces the hierarchical slow-fast structure of prior work with a Transformer that processes slow visual tokens and fast force tokens under a temporal causal mask, uses a GRU force encoder to preserve causality, and closes the force loop within an action chunk by cached deterministic DDIM denoising. To prevent modality collapse, the paper adds Virtual-target-based Representation Regularization (VRR): an auxiliary loss that predicts a virtual target computed from a quasi-static compliance model with an adaptive stiffness schedule. The method is evaluated on two real-robot tasks, box flipping and switch toggling, with 20 trials per condition, reporting success counts against Diffusion Policy, Reactive Diffusion Policy, and several ablations.

Significance. The architectural idea is coherent and timely: unifying slow vision and fast force via causal attention plus consistent cached inference is a plausible and potentially valuable simplification of the RDP hierarchy. If the empirical claims are confirmed, the paper would make a useful contribution to contact-rich imitation learning, and Algorithm 1 is concrete enough to be checked and extended. The switch-toggling result (18/20 vs 10/20 for RDP) is encouraging, and the failure-mode figures give qualitative support for the closed-loop narrative. However, the central claim of 'significantly outperforms' is not yet established by the reported experiments, and the reproducibility of the VRR objective is limited by unreported parameters. The paper is not internally inconsistent to the point of rejecting, but the evidence base must be substantially strengthened before the claims can be accepted.

major comments (4)
  1. [§IV-C, Table I, Abstract] The headline claim that ImplicitRDP 'significantly outperforms both vision-only and hierarchical baselines' is load-bearing and rests entirely on success counts over 20 trials with no confidence intervals, significance tests, repeated-seed variance, or trial logs. In Box Flipping, 18/20 vs RDP's 16/20 is not statistically significant (a two-sided Fisher exact test gives p≈0.34; the 95% Wilson intervals overlap substantially, roughly [0.70,0.97] vs [0.58,0.92]). The Switch Toggling comparison (18/20 vs 10/20) is more decisive, but the paper still does not report intervals or trial-level logs. The abstract's 'significantly' should be calibrated to the actual evidence, or the evidence must be expanded.
  2. [§III-C, Eq. (4)–(6), Table III] VRR is the second main contribution, but its formulation depends on several unreported free parameters: the ACP stiffness thresholds f_min, f_max and range k_min, k_max in Eq. (5), the orthogonal stiffness k_high used to assemble K, and the relative weighting of the virtual-target components in the augmented action. Since Table III shows a large gap between VRR (18/20) and raw force prediction (8/20 on box flipping), the reader cannot separate the effect of the proposed representation from tuning of these parameters. In addition, §III-C4 concedes that predicting x_vt is mathematically equivalent to predicting f_ext given x_real and K; the claimed benefits are therefore not additional information but a coordinate/weighting choice. This is not fatal, but the paper should report the stiffness values, provide a sensitivity analysis, and clarify how the adaptive weighting differs from a fixed
  3. [§IV-C, Table II; §III-B] Table II is the main evidence for the closed-loop SSL contribution, but it shows a task-dependent interaction that is not discussed. In Switch Toggling, the open-loop variant with VRR (ImplicitRDP w.o. SSL) achieves 15/20, only 3 trials below the full closed-loop model, whereas in Box Flipping the same variant is worse (4/20) than the open-loop variant without VRR (6/20). This means SSL is not uniformly 'critical' and that VRR can be unexpectedly effective even open-loop on one task. The paper should analyze this interaction and qualify the Q2 claim accordingly.
  4. [§III-D.3, §IV-A] The experimental comparison may be affected by hardware and controller choices that are not controlled across baselines. Section III-D.3 describes modifying the robot's impedance controller and tuning the PI integral gain k_i, and the paper uses a custom compliant fingertip. It is not stated whether DP and RDP were evaluated with the same fingertip, the same controller gains, and the same action-space conventions. Since the central claim is a comparison against baselines, these details are necessary for the result to be interpretable. The paper should also report training budgets, observation/action/execution horizons, latency steps l, and DDIM step counts.
minor comments (5)
  1. [§III-B, Alg. 1] The quantities h_o, h_a, h_e, and l are used but not defined numerically. Please also specify the actual control rate and how visual observations are sampled relative to the claimed 1–2 Hz slow modality when all data are recorded at 10 Hz.
  2. [§III-C, Eq. (5)] Please provide units for f_min, f_max, k_min, k_max and k_high, and clarify how the scalar k_adp is combined with the orthogonal high stiffness to form the anisotropic matrix K in Eq. (3).
  3. [§III-D.1] The statement that 'the three dimensions of Euler angles are independent' is not strictly correct for rotation representations. The intended point is probably that relative Euler increments avoid wrapping and Gimbal-lock discontinuities; please revise the wording.
  4. [§IV-C, Fig. 6] Please clarify how the attention weights in Fig. 6 are aggregated across episodes and whether the displayed curves are representative trials; as shown, they are difficult to interpret without error bands or multiple runs.
  5. [Availability] The manuscript promises 'code and videos' but includes no supplemental code or data release in the submission. Since the empirical claims are central, a reproducibility artifact or per-trial logs would greatly increase confidence.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an external empirical comparison; the only self-citation (RDP baseline) is not load-bearing, and the virtual-target/force equivalence is explicitly acknowledged.

full rationale

The central claim—that ImplicitRDP outperforms vision-only DP and hierarchical RDP on contact-rich tasks—is an empirical comparison supported by success counts in Tables I-IV, not a result derived from the paper's own assumptions. The virtual-target auxiliary objective is defined by Eq. (4) as a deterministic transform of force, and the paper itself states in Sec. III-C4: 'predicting x_vt is mathematically equivalent to predicting f_ext (given x_real and K)'. This is an explicit reparameterization with claimed learning advantages, not a hidden circular reduction, and the auxiliary outputs are discarded at inference. The only self-citation is the RDP baseline [6], whose authors overlap with the present paper; however, it is used as a comparison method with reported trials, not as a load-bearing premise justifying the method's correctness. No uniqueness theorem, ansatz-smuggled-via-citation, or fitted-parameter-as-prediction pattern is present. Statistical caveats about small trial counts concern the strength of the empirical evidence, not circularity. Therefore, no significant circularity is found; score 2 reflects only a minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced: the virtual target is a coordinate transformation of measured force, not a new object. The central method rests on the standard diffusion/DDIM framework, the quasi-static compliance assumption in Eq. 4, and the ACP adaptive-stiffness heuristic in Eq. 5. The paper does not report values for stiffness thresholds, horizons, or controller gains, so these hand-chosen constants are unstated free parameters. The main baseline RDP shares authors with this paper, but the comparison is external and task-defined.

free parameters (5)
  • f_min, f_max force thresholds
    Eq. 5; thresholds determining sensitivity to contact; values not reported; determine when VRR treats force as noise vs contact.
  • k_min, k_max stiffness range
    Eq. 5; bounds of adaptive stiffness; values not reported; scale K^{-1} and hence the magnitude of virtual-target deviation.
  • k_high orthogonal stiffness
    Sec. III-C2; high stiffness assigned to directions orthogonal to the force; value not reported.
  • Observation/action/execution horizons (h_o, h_a, h_e, l)
    Alg. 1; chunk lengths and latency are hyperparameters; values not reported but required to reproduce inference.
  • PI integral gain k_i
    Sec. III-D3; controller integral gain tuned so the robot tracks policy commands; value not reported; essential for the claimed reactivity.
axioms (6)
  • standard math Diffusion-policy conditional denoising and DDIM determinism (eta=0) are valid for action generation.
    Sec. III-A and III-B2; background from [1] and [28]; the paper assumes cached noise plus deterministic DDIM yields consistent closed-loop control.
  • domain assumption The compliance model in Eq. (3) and the quasi-static drop of M and D in Eq. (4) describe the contact mechanics.
    Sec. III-C1; the virtual target is derived from a spring-mass-damper system with inertia and damping ignored.
  • ad hoc to paper The ACP adaptive-stiffness assignment in Eq. (5) gives appropriate per-phase weighting.
    Sec. III-C2; heuristic from [9]; no validation that the k_adp curve is optimal or even necessary for these tasks.
  • domain assumption GRU force encoding plus a causal attention mask prevents future information leakage.
    Sec. III-B1; architecture premise for closed-loop causality; if leakage occurs, the reactive-control claim fails.
  • domain assumption The custom compliant fingertip and tuned PI controller produce force signals that reflect contact and are trackable.
    Secs. III-D2 and III-D3; hardware and controller modifications are part of the method and are not independently validated.
  • domain assumption The measured end-effector F/T readings are a faithful proxy for contact force.
    Sec. IV-A; a 6-axis F/T sensor is used as the contact-force signal, with no filtering or calibration details reported.

pith-pipeline@v1.3.0-alltime-deepseek · 10871 in / 16309 out tokens · 156401 ms · 2026-08-03T16:58:13.294760+00:00 · methodology

0 comments
read the original abstract

Human-level contact-rich manipulation relies on the distinct roles of two key modalities: vision provides spatially rich but temporally slow global context, while force sensing captures rapid local contact dynamics. Integrating these signals is challenging due to their fundamental frequency and informational disparities. In this work, we propose ImplicitRDP, a unified end-to-end visual-force diffusion policy that integrates visual planning and reactive force control within a single network. We introduce Structural Slow-Fast Learning, a mechanism utilizing causal attention to simultaneously process asynchronous visual and force tokens, allowing the policy to perform rapid force control at the action rate while maintaining the temporal coherence of action chunks. Furthermore, to mitigate modality collapse where end-to-end models fail to adjust the weights across different modalities, we propose Virtual-target-based Representation Regularization. This auxiliary objective maps force feedback into the same space as the action, providing a stronger, physics-grounded learning signal than raw force prediction. Extensive experiments on contact-rich tasks demonstrate that ImplicitRDP significantly outperforms both vision-only and hierarchical baselines, achieving superior reactivity and success rates with a streamlined training pipeline. Code and videos are available at https://implicit-rdp.github.io.

Figures

Figures reproduced from arXiv: 2512.10946 by Cewu Lu, Chuan Wen, Fangyuan Zhou, Han Xue, Jun Lv, Shirun Tang, Wendi Chen, Yang Jin, Yi Wang.

Figure 1
Figure 1. Figure 1: Structural Slow-Fast Learning: we leverage a temporally causal structure to enable end-to-end closed-loop force-based control within an action chunk. Virtual-target-based Representation Regularization: we further incorporate virtual target prediction as an auxiliary task to prevent modality collapse. processing force to predict reactive real actions in a closed loop. While intuitive, this explicit separati… view at source ↗
Figure 2
Figure 2. Figure 2: Network Architecture of ImplicitRDP. We enforce a temporally causal structure using a GRU for force signal encoding and a causal attention mask for action-force interaction, which enables structural slow￾fast learning. 1) Temporally Causal Structure: As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hardware Setup. The system utilizes a Flexiv Rizon 4s robot arm. A handle and button are mounted between the seventh joint and 6-axis F/T sensor for kinematic teaching. We also design a custom compliant fingertip to ensure distinctive reactivity signals during contact-rich interactions. A. Experimental Setup Our hardware setup (see [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Box Flipping Task and Failure Cases. The goal is to push a thin phone box against a fixture to flip it upright while maintaining a delicate force limit (< 14N). Vision-only or open-loop baselines lack closed-loop, force-based adjustment and apply excessive force, resulting in squeezing the fingertip. ImplicitRDP successfully utilizes the force feedback to complete the task safely. Keep Contact ImplicitRDP … view at source ↗
Figure 5
Figure 5. Figure 5: Switch Toggling Task and Failure Cases. The robot has to locate and apply a specific force to toggle a circuit breaker switch. DP tends to initiate the toggling motion prematurely before the triggering force threshold is reached, while RDP often misses the precise contact location due to latent compression errors. ImplicitRDP accurately approaches the switch and perceives force to toggle the switch success… view at source ↗
Figure 6
Figure 6. Figure 6: Attention Weight Visualization. We visualize the summed attention weights of visual tokens and force tokens from the first transformer layer in the switch toggling task. The weights are averaged over all heads and all queries. A sliding window of size 10 is applied to smooth the curves. approach phase. We hypothesize that this is because the fast policy in RDP compresses raw actions into a latent space, le… 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 3 Pith papers

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

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

    cs.RO 2026-07 conditional novelty 6.0

    Injecting recent end-effector force into a pretrained VLA through a zero-initialized reactive action expert plus online DAgger improves contact-rich manipulation over vision-only post-training.

  2. Seeing Touch from Motion: A Unified Modality-Aware Visuo-Tactile Policy with Tactile Motion Correlation

    cs.RO 2026-06 unverdicted novelty 5.0

    A visuo-tactile policy learning method that exploits tactile motion correlation for contact state distinction and Mixture-of-Transformers for cross-modal fusion.

  3. Learning Versatile Humanoid Manipulation with Touch Dreaming

    cs.RO 2026-04 conditional novelty 5.0

    HTD, a multimodal transformer policy trained with behavioral cloning and touch dreaming to predict future tactile latents, achieves a 90.9% relative success rate improvement over baselines on five real-world contact-r...

Reference graph

Works this paper leans on

33 extracted references · 17 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Diffusion policy: Vi- suomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burch- fiel, R. Tedrake, and S. Song, “Diffusion policy: Vi- suomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025. 1, 2

    2025

  2. [2]

    Learn- ing fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learn- ing fine-grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023. 1, 2

    Pith/arXiv arXiv 2023

  3. [3]

    π 0: A vision-language-action flow model for general robot control,

    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. 1, 2

    Pith/arXiv arXiv 2024

  4. [4]

    π 0.5: a vision-language-action model with open-world generalization,

    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. 1, 2

    Pith/arXiv arXiv 2025

  5. [5]

    Gen-0: Embodied foundation mod- els that scale with physical interaction,

    G. A. Team, “Gen-0: Embodied foundation mod- els that scale with physical interaction,”Generalist AI Blog, 2025, https://generalistai.com/blog/preview- uqlxvb-bb.html. 1, 2

    2025

  6. [6]

    Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipula- tion,

    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 manipula- tion,”arXiv preprint arXiv:2503.02881, 2025. 1, 2, 3

    Pith/arXiv arXiv 2025

  7. [7]

    Factr: Force-attending curriculum train- ing for contact-rich policy learning,

    J. J. Liu, Y . Li, K. Shaw, T. Tao, R. Salakhutdinov, and D. Pathak, “Factr: Force-attending curriculum train- ing for contact-rich policy learning,”arXiv preprint arXiv:2502.17432, 2025. 2

    Pith/arXiv arXiv 2025

  8. [8]

    Tacdiffusion: Force-domain diffusion policy for precise tactile ma- nipulation,

    Y . Wu, Z. Chen, F. Wu, L. Chen, L. Zhang, Z. Bing, A. Swikir, S. Haddadin, and A. Knoll, “Tacdiffusion: Force-domain diffusion policy for precise tactile ma- nipulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 11 831–11 837. 2

    2025

  9. [9]

    Adaptive compliance policy: Learning approximate compliance for diffusion guided control,

    Y . Hou, Z. Liu, C. Chi, E. Cousineau, N. Kup- puswamy, S. Feng, B. Burchfiel, and S. Song, “Adaptive compliance policy: Learning approximate compliance for diffusion guided control,” in2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 4829–4836. 2, 4

    2025

  10. [10]

    Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manip- ulation,

    W. Liu, J. Wang, Y . Wang, W. Wang, and C. Lu, “Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manip- ulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1105–1112. 2

    2025

  11. [11]

    Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation,

    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 Automation Letters, 2025. 2

    2025

  12. [12]

    Compliant residual dagger: Improving real-world contact-rich ma- nipulation with human corrections,

    X. Xu, Y . Hou, Z. Liu, and S. Song, “Compliant residual dagger: Improving real-world contact-rich ma- nipulation with human corrections,”arXiv preprint arXiv:2506.16685, 2025. 2

    arXiv 2025

  13. [13]

    Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,

    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,”arXiv preprint arXiv:2505.22159, 2025. 2

    arXiv 2025

  14. [14]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat,et al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023. 2

    Pith/arXiv arXiv 2023

  15. [15]

    Qwen3 tech- nical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv,et al., “Qwen3 tech- nical report,”arXiv preprint arXiv:2505.09388, 2025. 2

    Pith/arXiv arXiv 2025

  16. [16]

    From slow bidirec- tional to fast autoregressive video diffusion models,

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shechtman, and X. Huang, “From slow bidirec- tional to fast autoregressive video diffusion models,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 963–22 974. 2

    2025

  17. [17]

    Unleashing large-scale video generative pre-training for visual robot manipu- lation,

    H. Wu, Y . Jing, C. Cheang, G. Chen, J. Xu, X. Li, M. Liu, H. Li, and T. Kong, “Unleashing large-scale video generative pre-training for visual robot manipu- lation,”arXiv preprint arXiv:2312.13139, 2023. 2

    Pith/arXiv arXiv 2023

  18. [18]

    Predictive inverse dynamics models are scalable learners for robotic manipulation,

    Y . Tian, S. Yang, J. Zeng, P. Wang, D. Lin, H. Dong, and J. Pang, “Predictive inverse dynamics models are scalable learners for robotic manipulation,”arXiv preprint arXiv:2412.15109, 2024. 2

    Pith/arXiv arXiv 2024

  19. [19]

    Video prediction policy: A generalist robot policy with predictive vi- sual representations,

    Y . Hu, Y . Guo, P. Wang, X. Chen, Y .-J. Wang, J. Zhang, K. Sreenath, C. Lu, and J. Chen, “Video prediction policy: A generalist robot policy with predictive vi- sual representations,”arXiv preprint arXiv:2412.14803,

    Pith/arXiv arXiv

  20. [20]

    Unified video action model,

    S. Li, Y . Gao, D. Sadigh, and S. Song, “Unified video action model,”arXiv preprint arXiv:2503.00200, 2025. 2

    Pith/arXiv arXiv 2025

  21. [21]

    Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets,

    C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta, “Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets,”arXiv preprint arXiv:2504.02792, 2025. 2

    Pith/arXiv arXiv 2025

  22. [22]

    Ta-vla: Elucidating the design space of torque-aware vision-language-action models,

    Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao, “Ta-vla: Elucidating the design space of torque-aware vision-language-action models,” arXiv preprint arXiv:2509.07962, 2025. 2, 3

    Pith/arXiv arXiv 2025

  23. [23]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020. 3

    2020

  24. [24]

    Bayesian learning via stochastic gradient langevin dynamics,

    M. Welling and Y . W. Teh, “Bayesian learning via stochastic gradient langevin dynamics,” inProceedings of the 28th international conference on machine learn- ing (ICML-11), 2011, pp. 681–688. 3

    2011

  25. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural infor- mation processing systems, vol. 30, 2017. 3

    2017

  26. [26]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. 3

    2016

  27. [27]

    On the properties of neural machine trans- lation: Encoder-decoder approaches,

    K. Cho, B. Van Merri ¨enboer, D. Bahdanau, and Y . Ben- gio, “On the properties of neural machine trans- lation: Encoder-decoder approaches,”arXiv preprint arXiv:1409.1259, 2014. 3

    Pith/arXiv arXiv 2014

  28. [28]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502,

    Pith/arXiv arXiv 2010

  29. [29]

    Compliance and force control for com- puter controlled manipulators,

    M. T. Mason, “Compliance and force control for com- puter controlled manipulators,”IEEE Transactions on Systems, Man, and Cybernetics, vol. 11, no. 6, pp. 418– 432, 2007. 3

    2007

  30. [30]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,”arXiv preprint arXiv:2202.00512, 2022. 5

    Pith/arXiv arXiv 2022

  31. [31]

    On the continuity of rotation representations in neural networks,

    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, 2019, pp. 5745–5753. 5

    2019

  32. [32]

    Universal manip- ulation interface: In-the-wild robot teaching without in-the-wild robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manip- ulation interface: In-the-wild robot teaching without in-the-wild robots,”arXiv preprint arXiv:2402.10329,

    Pith/arXiv arXiv

  33. [33]

    https://www.flexiv.com/products/rizon, 2024. 5

    2024