Pith. sign in

REVIEW 3 major objections 7 minor 73 references

FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FlowRAM claims that making 3D perception shrink with the flow-matching trajectory yields state-of-the-art robot manipulation: 82.3% average success on RLBench and a 12.0-point gain on high-precision tasks.

desk verdict A genuinely new coupling of sampling radius to flow time, with solid RLBench numbers—but the headline 12-point gain is not directly ablated and the radius schedule is under-specified. read the letter →

arxiv 2506.16201 v1 pith:4AI5JKAP submitted 2025-06-19 cs.RO cs.CV

Sen Wang , Le Wang , Sanping Zhou , Jingyi Tian , Jiayi Li , Haowen Sun , Wei Tang This is my paper
classification cs.ROcs.CV
keywords roboticmanipulationflowmatchingconditionalstatespacemodelsdynamicradiusscheduleregion-awareperceptionRLBenchhigh-precision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

FlowRAM claims that a robot can learn precise, fast manipulation by tying the perception region to the flow-matching process itself: as the generated pose progresses from noise to target, the model narrows its 3D attention from the whole scene to the contact area. On RLBench, it reports an average success rate of 82.3% across 10 multi-task settings and 52.0% on 7 high-precision tasks, beating the previous best policy by 12.0 percentage points and generating actions in as few as 2 integration steps. This matters because diffusion-based policies are accurate but slow, and previous 3D methods either voxelize at fixed resolution or fail to focus on task-relevant geometry. FlowRAM's answer is to make where to look a function of flow time, so global context and fine detail are both available without extra inference cost.

What carries the argument

The load-bearing mechanism is the Dynamic Radius Schedule (DRS) coupled with conditional flow matching. DRS produces a sequence of region-aware masks M_i = {(p_i, r_i) | i = 0 ... N}, where p_i is the noise-perturbed end-effector position at flow time i and r_i = (1 - i)(r_0 - r_min) + r_min shrinks monotonically with i. Because p_i converges toward the target pose while r_i shrinks, the point-cloud encoder samples progressively finer geometry around the task-relevant contact region, so the generation trajectory doubles as a coarse-to-fine attention schedule. Conditional flow matching supplies the velocity field v_theta(x_t, t, C) that transports a Gaussian sample to a 6-DoF keyframe pose through Euler integration, and Mamba, a selective state space model with linear-time sequence processing, fuses geometric, semantic, instruction, and proprioceptive tokens.

What would settle it

Run the same FlowRAM configuration on the 7 high-precision tasks with DRS replaced by global farthest-point sampling while keeping CFM at 2 inference steps; the paper's mechanism predicts average success should fall (its multi-task ablation shows a 3.7-point drop at i=2). A complementary check is to record the distance between the region center p_i and the final target pose at each flow step; if early-step centers are frequently far off-target, the shrinking-radius schedule is concentrating resolution on the wrong geometry.

Watch

Extended reading notes

Core claim

FlowRAM's central claim is that the flow-matching trajectory itself can serve as an attention schedule for 3D perception. At each discretized flow step i, the model defines a region mask centered at the noise-perturbed end-effector position p_i, whose radius r_i decreases monotonically as i grows. This Dynamic Radius Schedule turns the generation process into a coarse-to-fine perception process: the policy first sees the whole workspace, then concentrates sampled points on the geometry around the predicted contact point just as the pose estimate approaches the target. The velocity field v_theta is trained by regressing x_1 - x_0 along linear interpolation paths, so at inference the action pose is produced by integrating the ODE in as few as 2 steps. On that basis the paper reports 82.3% average success on 10 RLBench tasks, 52.0% average success on 7 high-precision tasks (12.0 percentage points above the prior best), and successful real-world execution on 49 of 60 episodes across 6 tasks.

Load-bearing premise

The load-bearing premise is that the noise-perturbed end-effector position stays a reliable center for the task-relevant region at every flow step while the radius shrinks; if early flow steps land off-target, the model spends its resolution on the wrong geometry, and the paper does not report the exact starting and ending radii needed to check this.

Editorial extensions

If this is right

  • If the reported results hold, diffusion-based manipulation policies can be replaced by flow-matching policies that generate keyframe poses in 2-4 integration steps, which puts learned high-precision control inside a real-time feedback loop.
  • The Dynamic Radius Schedule should transfer to other tasks where contact, insertion, or alignment determines success, because it does not depend on task-specific object models and can use other radius functions such as cosine annealing.
  • Using a state space model for multimodal fusion means the added geometric tokens from region-aware sampling do not incur quadratic attention cost, so scaling to more cameras or denser point clouds is cheaper than with Transformer-based policies.
  • Since CFM outperformed DDIM at the same inference step counts, the choice of training objective, not just the region-aware sampling, is part of the reported speed-accuracy advantage.
  • The high-precision setting results (52.0% average, 12.0 points above the previous best) imply that the largest gains from region-aware perception appear exactly where global-voxel and multi-stage methods lose geometric detail.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: an untested consequence of DRS is that the schedule's center p_i could be wrong early in generation; a robust variant might replace the noisy position with a predicted region-of-interest heatmap, which would keep resolution on target even if the initial noise sample is far from the true pose.
  • Inference: because the radius function f(i) is a free design choice, the reported linear schedule may not be optimal; learning or per-task tuning of r_0 and r_min could push the high-precision numbers higher, but the paper does not report those values.
  • Inference: if CFM's few-step advantage generalizes, the same architecture could be trained with a rectified-flow objective toward one-step generation, making the policy effectively a deterministic map from observation to action while retaining multimodal diversity through the sampled noise.
  • Inference: the paper's own failure-case analysis attributes some umbrella-insertion failures to a simulator bug and multi-task trade-offs to uniform sampling; reweighting task sampling could improve the reported averages beyond the stated numbers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes FlowRAM, a keyframe-based robotic manipulation policy that couples conditional flow matching (CFM) for 6-DoF pose generation with a Mamba-based multimodal fusion module and a Dynamic Radius Schedule (DRS) that progressively narrows 3D point-cloud sampling around the noise-perturbed end-effector position during flow integration. On the RLBench benchmark, the authors report a state-of-the-art average success rate of 82.3% over 10 multi-task settings and 52.0% over 7 high-precision tasks, the latter exceeding the previous best by 12.0 percentage points. They further report strong few-step inference performance (74.2% with 2 integration steps) and real-world deployment results on 6 tasks.

Significance. If the empirical claims hold, the paper offers a practical and conceptually simple way to make generative flow steps region-aware, with a measurable speed advantage from few-step inference. The experimental comparison is broad, includes the standard RLBench benchmark and a real-robot evaluation, and the reported per-task numbers are internally consistent with the stated averages. The inclusion of training/inference pseudocode is also a strength. That said, the load-bearing causal claim—that DRS is responsible for the high-precision margin—is not directly tested on the high-precision benchmark, and the mathematical specification of the core DRS schedule is ambiguous as written. These issues must be resolved before the paper can be accepted.

major comments (3)
  1. [3.3, Eq. (7)] The dynamic radius schedule as written is not reproducible: r_i = (1 - i)*(r0 - rmin) + rmin with i in {0,...,N} gives r_1 = rmin and negative radii for all i > 1, which contradicts the stated monotonicity rmin <= r_i < r_{i-1} < ... < r_0. If i is intended to be normalized by the total number of steps, that normalization is omitted. Please restate the schedule (e.g., r_i = (1 - i/N)*(r0 - rmin) + rmin) and report the exact values of r0 and rmin, including the values used for the 2-step and 50-step evaluations.
  2. [4.2-4.3, Tables 2, 3, C] The headline 12.0 percentage-point improvement on the 7 high-precision tasks is attributed to the Dynamic Radius Schedule, but the DRS ablation is only conducted on the 10-task multi-task benchmark. Table C shows DRS helps in that setting, yet there is no w/o DRS row for Table 2. Because DRS is the paper's main novelty and the high-precision benchmark is its main claim, the causal role of DRS in high-precision performance remains unsupported. Please add DRS ablations on the same 7-task high-precision benchmark.
  3. [3.3 and Table 4] The DRS mechanism centers sampling on the noise-perturbed position p_i. Since p_i = t*x1 + (1 - t)*x0, the target x1 lies within radius r_i only when r_i >= (1 - t)*||x0 - x1||, but the fixed schedule is independent of ||x0 - x1|| and no sensitivity to r0/rmin is reported. At the 2-step Euler integration used in Table 4, the only intermediate state is t = 0.5; if the radius has already shrunk below the typical displacement, the sampled geometric tokens may exclude the target, and the reported 2-step success could rely on the global RGB branch rather than the proposed region-aware mechanism. Please report the radius schedule values and include a sensitivity study or a discussion of this failure mode.
minor comments (7)
  1. [3.3 and Table B] The notation N1/N2 for sampled tokens is swapped between the text and Table B. In Section 3.3, F_geo is in R^{N1 x C} and F_rgb is in R^{N2 x C}, so N1 is the geometric token count and N2 the semantic token count; Table B instead lists 'sampled semantic tokens N1=4096' and 'sampled geometric tokens N2=1024'. Please align the notation.
  2. [3.2, Eq. (6)] The action representation declares arot in SO(3) but the next sentence states that a 6D rotation representation is used; please write arot in R^6 to avoid the contradiction.
  3. [3.3 and Fig. 2] The number of Mamba blocks N is a key architectural hyperparameter but it is not reported in Table B or anywhere else; please list its value.
  4. [Table 3] The checkmark/box symbols used to indicate DRS, P.E., and T.O. are not legible in the compiled text; please use explicit 'Yes/No' columns.
  5. [Table 1] The column header 'Slide Bolck' should read 'Slide Block'.
  6. [Abstract and Table 5] The claim of generating physically plausible actions 'in less than 4 time steps' for real-world tasks is not supported by a reported inference-step count in the real-robot experiments; please report the number of integration steps used in Table 5 or qualify the claim.
  7. [Section 2.2] The sentence beginning 'Where flow matching...' should begin with 'While' to read correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical benchmark comparisons, not derivations that reduce to their inputs.

full rationale

I walked the claimed derivation chain: the Dynamic Radius Schedule is defined as a design choice (Eq. 7) that centers a perception region on the noised pose and shrinks with a hand-specified schedule; the conditional flow matching objective (Eq. 11) is the standard flow matching loss regressing a velocity field; and the inference rule (Eq. 13) is ordinary Euler integration. None of these equations takes the reported success rates as an input, and neither the 82.3% multi-task average nor the 52.0% high-precision average is produced by fitting a parameter to the target result. The results are empirical comparisons against external RLBench baselines. The only self-citation, reference [64], appears in the related-work discussion of deterministic policies and is not load-bearing. The under-specification of r0 and rmin and the apparent missing normalization in Eq. 7 are reproducibility weaknesses, not circularity. Therefore I find no circular step.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

FlowRAM relies on standard flow matching and SSM mathematics, plus domain assumptions about keyframes, planner execution, and the DRS convergence behavior. The only assumption unique to this paper is that the noised pose is a reliable region center, which is not independently validated and depends on unreported radius values.

free parameters (4)
  • DRS radii r0 and rmin = not reported
    The linear radius schedule ri = (1 - i/N) * (r0 - rmin) + rmin in Section 3.3 depends on these concrete values; without them the core novelty cannot be reproduced exactly.
  • Loss weights lambda_CFM and lambda_open = 100 and 10
    Chosen by hand in Table B; they determine the balance between pose velocity regression and gripper state prediction.
  • Number of Mamba blocks N = not reported
    The architecture figures say 'Mamba Block x N' but Table B does not list N, so the fusion model cannot be rebuilt exactly.
  • Sampled token counts N1 and N2 = 4096 and 1024, with labels swapped in Table B
    These control the density of point-cloud and image tokens. Section 3.3 and Table B disagree on which count is geometric and which is semantic, creating a replication ambiguity.
assumptions (5)
  • standard math Conditional flow matching with linear interpolation xt = t*x1 + (1 - t)*x0 gives a tractable velocity-field objective (Eq. 2 and Eq. 10).
    Background assumption taken from cited flow matching works [38, 40, 41]; not proved in this paper but standard in the generative modeling literature.
  • standard math State-space models discretized with zero-order hold preserve long-range dependencies and linear complexity (Eq. 4).
    Background assumption from [22]; the paper relies on this for the efficiency claim.
  • domain assumption Keyframe-based manipulation reduces the problem to predicting the next bottleneck action pose, with keyframes defined by near-zero joint velocities and unchanged gripper state.
    Section 3.2 follows prior keyframe pipelines; if keyframe extraction is imperfect, the reported success rates conflate perception and planning errors.
  • domain assumption The learned velocity field trained with Logit-Normal time sampling transfers to fewer inference steps without significant distribution shift.
    Table 4 shows this empirically for the tested tasks, but the paper provides no theoretical guarantee that 2-step Euler integration matches 50-step training paths.
  • ad hoc to paper The noise-perturbed position p_i is a reliable center for the task-relevant region as the flow progresses.
    Section 3.3, Eq. (7) defines DRS around p_i. If early flow predictions are off-target, the shrinking radius would focus on the wrong location; this failure mode is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation." pith.science (2026). https://pith.science/paper/4AI5JKAP

@misc{pith2026250616201,
  author       = {Pith},
  title        = {Pith review of: FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AI5JKAP}},
  note         = {Machine review of arXiv:2506.16201}
}
read the original abstract

Robotic manipulation in high-precision tasks is essential for numerous industrial and real-world applications where accuracy and speed are required. Yet current diffusion-based policy learning methods generally suffer from low computational efficiency due to the iterative denoising process during inference. Moreover, these methods do not fully explore the potential of generative models for enhancing information exploration in 3D environments. In response, we propose FlowRAM, a novel framework that leverages generative models to achieve region-aware perception, enabling efficient multimodal information processing. Specifically, we devise a Dynamic Radius Schedule, which allows adaptive perception, facilitating transitions from global scene comprehension to fine-grained geometric details. Furthermore, we integrate state space models to integrate multimodal information, while preserving linear computational complexity. In addition, we employ conditional flow matching to learn action poses by regressing deterministic vector fields, simplifying the learning process while maintaining performance. We verify the effectiveness of the FlowRAM in the RLBench, an established manipulation benchmark, and achieve state-of-the-art performance. The results demonstrate that FlowRAM achieves a remarkable improvement, particularly in high-precision tasks, where it outperforms previous methods by 12.0% in average success rate. Additionally, FlowRAM is able to generate physically plausible actions for a variety of real-world tasks in less than 4 time steps, significantly increasing inference speed.

Figures

Figures reproduced from arXiv: 2506.16201 by the authors.

Figure 1
Figure 1. The diagram of the proposed FlowRAM. Consider the insert USB in computer task. Previous arts have at￾tempted to capture information about the geometric details of USB and ports, but fail severely due to lacking task-relevant perception. In contrast, FlowRAM equipped with a dynamic radius schedule successfully preserves these geometric details and ensures accu￾rate alignment of the USB and port. The pi, ri in the fig… view at source ↗
Figure 2
Figure 2. The overall framework of FlowRAM. FlowRAM is a Mamba-based framework, where multimodal Mamba processes multi￾view RGB images, geometric information, instructions, and robot proprioception. In flow matching, noise-perturbed actions are transported to target actions guided by observations. During training, a dynamic radius schedule adjusts the perception radius over random discrete time steps, enabling the model to ca… view at source ↗
Figure 3
Figure 3. Overview of the tasks. Six high-precision tasks from RLBench and two real-world tasks are visualized. Notably, the Unplug Charger is the inverse process of the Plug Charger in simulator. the destination keyframe action by integrating the flow from t = 0 to t = 1 over several steps. Formally, the inference process is defined as follows: xt+∆t = xt + vθ(xt, t, C)∆t, (13) The x1 after integration represents the predict… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance of CFM in High-Precision Tasks. We compare FlowRAM with DDPM and DDIM across 7 high-precision tasks, results demonstrate that CFM achieves strong performance in just 2 steps, showing excellent results even in high-precision scenarios. i-steps 2 4 8 16 32 DD…
Figure 5
Figure 5. Figure 5: Perception Region Illustration. We visualize different regions controlled by the dynamic radius schedule in real-world. trajectories by executing human-defined waypoints. Images are initially captured at a resolution of 2048 × 1536 and downsampled to 256 × 192. For inf…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 54 canonical work pages

  1. [1]

    Riemannian flow matching policy for robot motion learning

    Max Braun, No ´emie Jaquier, Leonel Rozo, and Tamim As- four. Riemannian flow matching policy for robot motion learning. arXiv preprint arXiv:2403.10672, 2024. 3

    arXiv 2024

  2. [2]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2

    arXiv 2022

  3. [3]

    GenCHiP: Generating Robot Policy Code for High-Precision and Contact-Rich Manipulation Tasks

    Kaylee Burns, Ajinkya Jain, Keegan Go, Fei Xia, Michael Stark, Stefan Schaal, and Karol Hausman. Genchip: Gen- erating robot policy code for high-precision and contact-rich manipulation tasks. arXiv preprint arXiv:2404.06645, 2024. 1

    work page Pith review arXiv 2024

  4. [4]

    Mamba policy: Towards efficient 3d diffusion policy with hybrid selective state models

    Jiahang Cao, Qiang Zhang, Jingkai Sun, Jiaxu Wang, Hao Cheng, Yulin Li, Jun Ma, Yecheng Shao, Wen Zhao, Gang Han, et al. Mamba policy: Towards efficient 3d diffusion policy with hybrid selective state models. arXiv preprint arXiv:2409.07163, 2024. 3

    arXiv 2024

  5. [5]

    Scanrefer: 3d object localization in rgb-d scans using natural language

    Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202–221. Springer, 2020. 1

    work page 2020

  6. [6]

    Flow matching on gen- eral geometries

    Ricky TQ Chen and Yaron Lipman. Flow matching on gen- eral geometries. In ICLR, 2024. 2

    work page 2024

  7. [7]

    Polarnet: 3d point clouds for language-guided robotic manipulation

    Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. CoRL, 2023. 2

    work page 2023

  8. [8]

    Confidence-based policy learning from demonstration using gaussian mixture models

    Sonia Chernova and Manuela Veloso. Confidence-based policy learning from demonstration using gaussian mixture models. In Proceedings of the 6th international joint confer- ence on Autonomous agents and multiagent systems , pages 1–8, 2007. 2

    work page 2007

Show all 73 references
  1. [9]

    Diffusion policy: Visuomotor policy learning via action dif- fusion

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action dif- fusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 3

    2023

  2. [10]

    Learn- ing robotic manipulation policies from point clouds with conditional flow matching

    Eugenio Chisari, Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox, and Abhinav Valada. Learn- ing robotic manipulation policies from point clouds with conditional flow matching. In CoRL, 2024. 3

    2024

  3. [11]

    Reducing the barrier to entry of complex robotic software: a moveit! case study

    David Coleman, Ioan Sucan, Sachin Chitta, and Niko- laus Correll. Reducing the barrier to entry of complex robotic software: a moveit! case study. arXiv preprint arXiv:1404.3785, 2014. 8

    2014 arXiv

  4. [12]

    Flow matching in latent space

    Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698,

    arXiv

  5. [13]

    Efficient video prediction via sparsely conditioned flow matching

    Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Efficient video prediction via sparsely conditioned flow matching. In ICCV, pages 23263–23274, 2023. 3

    2023

  6. [14]

    Hamba: Single-view 3d hand reconstruction with graph-guided bi-scanning mamba,

    Haoye Dong, Aviral Chharia, Wenbo Gou, Francisco Vicente Carrasco, and Fernando De la Torre. Hamba: Single-view 3d hand reconstruction with graph-guided bi-scanning mamba,

  7. [15]

    Palm-e: An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. In ICML, pages 8469–8488. PMLR, 2023. 2

    2023

  8. [16]

    Deep se (3)-equivariant geometric reasoning for precise placement tasks

    Ben Eisner, Yi Yang, Todor Davchev, Mel Vecerik, Jonathan Scholz, and David Held. Deep se (3)-equivariant geometric reasoning for precise placement tasks. In ICLR, 2024. 1

    2024

  9. [17]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 3, 5

    2024

  10. [18]

    Actionflow: Equivari- ant, accurate, and efficient policies with spatially symmetric flow matching

    Niklas Funk, Julen Urain, Joao Carvalho, Vignesh Prasad, Georgia Chalvatzaki, and Jan Peters. Actionflow: Equivari- ant, accurate, and efficient policies with spatially symmetric flow matching. arXiv preprint arXiv:2409.04576, 2024. 3

    2024 arXiv

  11. [19]

    Act3d: 3d feature field transformers for multi-task robotic manipulation

    Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Ka- terina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. In CoRL, 2023. 1, 2, 6, 7

    2023

  12. [20]

    Rvt: Robotic view transformer for 3d object manipulation

    Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In CoRL, pages 694–710. PMLR, 2023. 2, 3

    2023

  13. [21]

    Rvt2: Learning precise manipulation from few demonstrations

    Ankit Goyal, Valts Blukis, Jie Xu, Yijie Guo, Yu-Wei Chao, and Dieter Fox. Rvt2: Learning precise manipulation from few demonstrations. RSS, 2024. 1, 2, 3, 6, 7

    2024

  14. [22]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 5

    2023 arXiv

  15. [23]

    Efficiently mod- eling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R´e. Efficiently mod- eling long sequences with structured state spaces. In ICLR,

  16. [24]

    Instruction-driven history-aware policies for robotic manip- ulations

    Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manip- ulations. In CoRL, pages 175–187. PMLR, 2023. 2, 7, 1

    2023

  17. [25]

    Mambair: A simple baseline for image restoration with state-space model

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 3

    2024

  18. [26]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. NeurIPS, 33:6840–6851, 2020. 7

    2020

  19. [27]

    3d-llm: Inject- ing the 3d world into large language models

    Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Inject- ing the 3d world into large language models. NeurIPS, 36: 20482–20494, 2023. 1

    2023

  20. [28]

    Q-attention: En- abling efficient learning for vision-based robotic manipula- tion

    Stephen James and Andrew J Davison. Q-attention: En- abling efficient learning for vision-based robotic manipula- tion. IEEE Robotics and Automation Letters , 7(2):1612– 1619, 2022. 4

    2022

  21. [29]

    Stephen James, Zicong Ma, David Rovick Arrojo, and An- drew J. Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Let- ters, 2020. 2, 6, 1

    2020

  22. [30]

    Coarse-to-fine q-attention: Efficient learn- ing for visual robotic manipulation via discretisation

    Stephen James, Kentaro Wada, Tristan Laidlow, and An- drew J Davison. Coarse-to-fine q-attention: Efficient learn- ing for visual robotic manipulation via discretisation. In CVPR, pages 13739–13748, 2022. 2, 4

    2022

  23. [31]

    Bc-z: Zero-shot task generalization with robotic imitation learning

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Fred- erik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991–1002. PMLR, 2022. 2

    2022

  24. [32]

    Coarse-to-fine imitation learning: Robot ma- nipulation from a single demonstration

    Edward Johns. Coarse-to-fine imitation learning: Robot ma- nipulation from a single demonstration. In ICRA, pages 4613–4619. IEEE, 2021. 4

    2021

  25. [33]

    3d diffuser actor: Policy diffusion with 3d scene rep- resentations

    Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragki- adaki. 3d diffuser actor: Policy diffusion with 3d scene rep- resentations. CoRL, 2024. 3, 4, 6, 7

    2024

  26. [34]

    Diff-bgm: A diffusion model for video background mu- sic generation

    Sizhe Li, Yiming Qin, Minghang Zheng, Xin Jin, and Yang Liu. Diff-bgm: A diffusion model for video background mu- sic generation. In CVPR, pages 27348–27357, 2024. 3

    2024

  27. [35]

    Pointmamba: A simple state space model for point cloud analysis

    Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 2, 3, 5, 7

    2024

  28. [36]

    MTMamba: Enhancing multi-task dense scene understanding by mamba-based de- coders

    Baijiong Lin, Weisen Jiang, Pengguang Chen, Yu Zhang, Shu Liu, and Ying-Cong Chen. MTMamba: Enhancing multi-task dense scene understanding by mamba-based de- coders. In ECCV, 2024. 3

    2024

  29. [37]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Doll ´ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117–2125,

  30. [38]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3

    2023

  31. [39]

    Robomamba: Mul- timodal state space model for efficient robot reasoning and manipulation

    Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Mul- timodal state space model for efficient robot reasoning and manipulation. NeurIPS, 2024. 3

    2024

  32. [40]

    Rectified flow: A marginal preserving approach to optimal transport

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022. 2, 3

    2022 arXiv

  33. [41]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3

    2023

  34. [42]

    Vmamba: Visual state space model

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166,

    arXiv

  35. [43]

    Pose-and-shear-based tactile servoing

    John Lloyd and Nathan F Lepora. Pose-and-shear-based tactile servoing. The International Journal of Robotics Re- search, 43(7):1024–1055, 2024. 1

    2024

  36. [44]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  37. [45]

    Manigaussian: Dynamic gaus- sian splatting for multi-task robotic manipulation

    Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Ji- wen Lu, and Yansong Tang. Manigaussian: Dynamic gaus- sian splatting for multi-task robotic manipulation. arXiv preprint arXiv:2403.08321, 2024. 2, 6, 7

    2024 arXiv

  38. [46]

    Madiff: Motion-aware mamba diffusion models for hand trajectory prediction on egocentric videos, 2024

    Junyi Ma, Xieyuanli Chen, Wentao Bao, Jingyi Xu, and Hes- heng Wang. Madiff: Motion-aware mamba diffusion models for hand trajectory prediction on egocentric videos, 2024. 3

    2024

  39. [47]

    Hierarchical diffusion policy for kinematics-aware multi- task robotic manipulation

    Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multi- task robotic manipulation. In CVPR, pages 18081–18090,

  40. [48]

    Iris: Implicit reinforcement without interaction at scale for learn- ing control from offline robot manipulation data

    Ajay Mandlekar, Fabio Ramos, Byron Boots, Silvio Savarese, Li Fei-Fei, Animesh Garg, and Dieter Fox. Iris: Implicit reinforcement without interaction at scale for learn- ing control from offline robot manipulation data. In ICRA, pages 4414–4420. IEEE, 2020. 2

    2020

  41. [49]

    S4nd: Modeling images and videos as multidimensional signals with state spaces

    Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher R ´e. S4nd: Modeling images and videos as multidimensional signals with state spaces. NeurIPS, 35:2846–2861, 2022. 3

    2022

  42. [50]

    Read: Retrieval-enhanced asymmetric diffusion for motion plan- ning

    Takeru Oba, Matthew Walter, and Norimichi Ukita. Read: Retrieval-enhanced asymmetric diffusion for motion plan- ning. In CVPR, pages 17974–17984, 2024. 3

    2024

  43. [51]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195–4205, 2023. 5

    2023

  44. [52]

    Alvinn: An autonomous land vehicle in a neural network

    Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. NeurIPS, 1, 1988. 4

    1988

  45. [53]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652–660, 2017. 5

    2017

  46. [54]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, pages 8748–8763. PMLR, 2021. 5

    2021

  47. [55]

    A generalist agent

    Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Sprin- genberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 2

    2022 arXiv

  48. [56]

    V-rep: A versatile and scalable robot simulation framework

    Eric Rohmer, Surya PN Singh, and Marc Freese. V-rep: A versatile and scalable robot simulation framework. In IROS, pages 1321–1326. IEEE, 2013. 6

    2013

  49. [57]

    Learning from demonstration

    Stefan Schaal. Learning from demonstration. NeurIPS, 9,

  50. [58]

    Cliport: What and where pathways for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipulation. In CoRL, pages 894–906. PMLR, 2022. 2

    2022

  51. [59]

    Perceiver- actor: A multi-task transformer for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver- actor: A multi-task transformer for robotic manipulation. In CoRL, pages 785–799. PMLR, 2023. 2, 4, 6, 7, 1, 3

    2023

  52. [60]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 7

    2010 arXiv

  53. [61]

    Octo: An open-source generalist robot policy

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 4

    2024 arXiv

  54. [62]

    Deep generative models in robotics: A survey on learning from multimodal demonstrations

    Julen Urain, Ajay Mandlekar, Yilun Du, Mahi Shafiullah, Danfei Xu, Katerina Fragkiadaki, Georgia Chalvatzaki, and Jan Peters. Deep generative models in robotics: A survey on learning from multimodal demonstrations. arXiv preprint arXiv:2408.04380, 2024. 2

    2024 arXiv

  55. [63]

    Unleashing large-scale video generative pre-training for visual robot manipulation

    Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In ICLR, 2024. 4

    2024

  56. [64]

    Learning to refactor action and co-occurrence fea- tures for temporal action localization

    Kun Xia, Le Wang, Sanping Zhou, Nanning Zheng, and Wei Tang. Learning to refactor action and co-occurrence fea- tures for temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884–13893, 2022. 2

    2022

  57. [65]

    Chaineddiffuser: Uni- fying trajectory diffusion and keypose prediction for robotic manipulation

    Zhou Xian, Nikolaos Gkanatsios, Theophile Gervet, Tsung- Wei Ke, and Katerina Fragkiadaki. Chaineddiffuser: Uni- fying trajectory diffusion and keypose prediction for robotic manipulation. In CoRL, 2023. 2

    2023

  58. [66]

    M2t2: Multi-task masked transformer for object-centric pick and place

    Wentao Yuan, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. M2t2: Multi-task masked transformer for object-centric pick and place. In CoRL, pages 3619–3630. PMLR, 2023. 2

    2023

  59. [67]

    Gnfactor: Multi-task real robot learning with generalizable neural feature fields

    Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. InCoRL, pages 284–

  60. [68]

    3d diffusion policy: Gen- eralizable visuomotor policy learning via simple 3d repre- sentations

    Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Gen- eralizable visuomotor policy learning via simple 3d repre- sentations. In RSS, 2024. 1, 2, 3

    2024

  61. [69]

    Affordance-based robot manipulation with flow matching

    Fan Zhang and Michael Gienger. Affordance-based robot manipulation with flow matching. arXiv preprint arXiv:2409.01083, 2024. 3

    2024

  62. [70]

    Point could mamba: Point cloud learning via state space model

    Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 3

    2024 arXiv

  63. [71]

    Learning fine-grained bimanual manipulation with low-cost hardware

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In RSS, 2023. 1

    2023

  64. [72]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, pages 5745–5753, 2019. 4

    2019

  65. [73]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 3 FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robo...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.