Pith. sign in

REVIEW 4 major objections 6 minor 21 references

Latent-space steering of a frozen flow policy improves whole-body mobile manipulation beyond imitation-only training.

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 15:20 UTC pith:GLAUJQJW

load-bearing objection A credible engineering extension of DSRL to flow-based whole-body policies, with a real latent-drift question and too little statistical evidence to fully trust the headline numbers. the 4 major comments →

arxiv 2607.26460 v1 pith:GLAUJQJW submitted 2026-07-29 cs.RO

RLMM-Flow: A Flow-based Mobile Manipulation Framework with Latent-Space Reinforcement Learning

Shuhang Wang , Ziming Li , Hui Cheng This is my paper
classification cs.RO
keywords mobile manipulationflow matching policieslatent-space reinforcement learningoffline RL post-trainingwhole-body motion generationlatent steeringaction critic warm-upcoarse-to-fine latent steering
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.

The paper sets out to show that a flow-based generative policy for whole-body mobile manipulation, trained by imitation on expert demonstrations, can be improved by freezing it and steering its initial noise in latent space with offline reinforcement learning, rather than updating the policy directly in action space. It argues that action-space RL post-training distorts the learned motion prior, while latent-space steering preserves that prior and pushes generated trajectory chunks toward higher reward. To make latent optimization tractable for high-dimensional whole-body chunks, the paper introduces two stabilizers: warming up the action critic before training the latent critic, and a coarse-to-fine schedule that first optimizes a horizon-shared latent and later activates time-dependent residual corrections. If the paper is right, the result is higher success, fewer collisions, and smoother trajectories than imitation-only flow policies and existing RL post-training baselines, while keeping flow-based inference fast enough for real-time deployment.

Core claim

The central claim is that a pretrained flow policy for mobile manipulation can be treated as a frozen generator and improved by optimizing a latent steering network that re-parameterizes the initial noise fed into the flow. The action critic QA evaluates decoded whole-body action chunks; the latent critic QW is trained to distill QA's values onto latent noise samples; and the latent actor maximizes QW through reparameterized sampling. Two stabilizers make this work: QA is warmed up alone for 20,000 updates before QW and the actor are trained, giving the latent critic a stable target; and coarse-to-fine steering first uses a horizon-pooled, repeated latent (effective dimension 9 instead of 50

What carries the argument

The load-bearing mechanism is the frozen flow policy combined with a cascaded latent steering network. The steering network predicts base-related latent components before arm-related components, then applies a coarse-to-fine transformation: average pooling along the horizon compresses the latent to a horizon-shared vector, which is repeated back to the original shape; after a switch step, a residual branch adds time-dependent corrections with a scale factor. The latent critic QW provides the training signal by distilling values from the action critic QA on actions decoded by the frozen flow policy, enabling gradients to flow through the sampled latent without backpropagating through the flow

Load-bearing premise

The method assumes that increasing the latent critic's estimated value for a noise vector reliably increases the quality of the action chunk that the frozen flow policy decodes from that noise; if the flow transport is not smooth or the latent critic is inaccurate off the expert manifold, steering can push the policy away from the learned motion prior.

What would settle it

Decode a set of noise vectors with the frozen flow policy, compute the latent critic's predicted value and the actual reward of each decoded action chunk (or rollout), and check whether high latent-predicted values correspond to high measured rewards. A weak or negative rank correlation would show that latent value maximization does not transfer to decoded action quality, falsifying the core mechanism.

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

If this is right

  • Latent-space steering can improve imitation-learned generative policies without modifying policy weights or backpropagating through the full flow transport, so inference time stays close to the frozen policy's cost.
  • Warming up the action critic before latent distillation stabilizes offline RL for high-dimensional action chunks, a recipe that could transfer to other generative policy classes.
  • Coarse-to-fine latent search reduces effective dimensionality early in training, suggesting a general principle for latent-space RL with long-horizon action chunks.
  • Offline RL post-training can exploit mixed-quality datasets that include structured suboptimal trajectories, broadening value coverage beyond the expert manifold.
  • The post-trained policy transfers zero-shot from simulation to real hardware without real-world fine-tuning.

Where Pith is reading between the lines

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

  • Inference: If the core mechanism holds, the initial noise of any flow or diffusion policy can serve as a compact, steerable policy parameterization, so the approach may extend to other multi-step generators where full backpropagation is expensive.
  • Inference: The cascaded base-before-arm latent prediction encodes a useful kinematic dependency (global navigation before local manipulation) that could be tested on other heterogeneous action spaces, such as legged locomotion with an arm.
  • Inference: A direct test of the paper's assumption would be to rank noise samples by the latent critic's value and compare against actual decoded trajectory rewards; a weak correlation would indicate the residual schedule and warm-up are compensating for an uninformative value landscape.
  • Inference: The deliberate use of relaxed-planner suboptimal trajectories is a cheap data source for offline RL; the sensitivity of the method to the high-quality-to-suboptimal mixture ratio is an obvious experimental extension not explored in the paper.

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 / 6 minor

Summary. The paper proposes RLMM-Flow, a two-stage offline RL post-training method for flow-based whole-body mobile manipulation policies. Stage 1 pretrains a point-cloud-conditioned rectified flow policy on expert whole-body action chunks. Stage 2 freezes the flow policy and trains a Gaussian latent steering network, guided by a distilled latent critic, to choose initial noise that maximizes a trajectory-level reward. To stabilize high-dimensional latent optimization, the action critic is warmed up before latent training, and latent steering proceeds coarse-to-fine: a horizon-pooled latent first, then a scaled temporal residual. Evaluations on simulation (seen/unseen) and real-world deployment report improvements over imitation-only flow/diffusion baselines and DSRL-NA, plus ablations of the two proposed components.

Significance. If the results hold, the paper makes a useful empirical contribution: it demonstrates that latent-space steering of a frozen flow policy can serve as an offline post-training mechanism for whole-body mobile manipulation, and it proposes two plausible stabilization techniques (action-critic warm-up and coarse-to-fine latent steering). The reward formulation, dataset construction, and zero-shot real-world transfer are described in unusual detail, and there is no circularity between the derivation and the empirical claims. The main weaknesses are that all quantitative evidence is single-run and without variance, and the latent actor is not regularized toward the flow prior, leaving a plausible critic-exploitation failure mode. The significance is therefore conditional on additional robustness evidence.

major comments (4)
  1. [Tables 1-2, Figures 6-7 (Simulation Results, Real-world Deployment)] All quantitative comparisons are single point estimates. No standard deviations, confidence intervals, number of seeds, or significance tests are reported. The headline gains over DSRL-NA are modest (Table 1: 40→46 seen Succ., 30→38 unseen; Table 2: 24→32 real-world Succ.), so without variance information it is impossible to know whether the proposed components are real improvements or noise. Because the central claim is empirical superiority, please report multiple seeds (or confidence intervals) for all tables and reward curves, and state the number of seeds used.
  2. [Algorithm 1, Eqs. (37)-(39); Eq. (7)] The latent actor maximizes Q_W(s, w_π) with no KL penalty, trust region, or prior-distance constraint. The coarse branch in Eq. (7) uses w_coarse = Repeat(AvgPool_H(w_raw)), so the actor's mean μ_W can move arbitrarily far from the flow policy's training prior N(0,I); λ_res only scales the residual after t_cf and does not bound the coarse component. Q_W is trained on q_t(w|s) = ρN(0,I) + (1−ρ)sg[π_W^CTF,t(w|s)], so if π_W drifts into low-density latent regions, Q_W can extrapolate and the actor can exploit overestimation. The paper reports no measurement of latent drift or of Q_W accuracy as a function of distance from the prior. Please add a proximal/prior penalty or trust region, and report the latent distribution distance before/after training, or otherwise show that drift is bounded.
  3. [Eqs. (2), (24)-(26) vs. Evaluation Metrics] The reward is computed on the predicted action chunk τ (collision-sphere positions from forward kinematics of the chunk), while success and collision metrics are measured on the executed trajectory under the shared MPC controller. The paper does not report MPC tracking error or demonstrate that optimizing the planned reward improves the executed trajectory. Since the abstract's claim concerns realized task success, the connection between training signal and reported metrics needs evidence. Please quantify tracking error or evaluate the reward on the executed trajectories.
  4. [Eqs. (37)-(39); assumption on latent-value transfer] The method assumes that increasing the distilled latent critic Q_W is predictive of improved decoded action quality from the frozen flow transport. If the transport is non-smooth or Q_W is uninformative off the expert manifold, steering can move away from the learned prior. The paper provides only aggregate reward curves for this. Please give direct evidence, e.g., correlation/calibration plots of Q_W versus Q_A on held-out latents, or an analysis of Q_W's landscape along the actor update direction.
minor comments (6)
  1. [Abstract] There are missing spaces in the abstract text: 'aflow-basedmobilemanipulationframeworkthatcom-bines' and similar. Please copyedit.
  2. [Eq. (34)] The paper says 'log-normal noise-level sampling' but Eq. (34) uses a logit-normal schedule (z~N(-1.2,1.2^2), t=sigmoid(z)). Please use consistent terminology.
  3. [Algorithm 1 / Eq. (37)] The stop-gradient notation sg[·] is used before being defined. Add a definition at first use.
  4. [Table 5] The entry 'Actor distribution – Diagonal Gaussian' is a model choice rather than a hyperparameter value; clarify its status.
  5. [Related Work and Baseline] The relation between 'DSRL' and 'DSRL-NA' is not fully explicit. Clarify whether DSRL-NA is the exact configuration of Wagenmaker et al. or a reimplementation, and state whether its latent dimensionality was tuned.
  6. [Success criterion] The success criterion includes 'neither collisions nor dynamic-constraint violations' but does not mention joint-limit violations; Joint Viol. is reported separately. State explicitly whether a rollout with a joint-limit violation can still be counted as successful.

Circularity Check

0 steps flagged

No circularity: empirical RL post-training; QW→QA distillation is an explicit training mechanism and not used as an evaluation prediction.

full rationale

RLMM-Flow is an empirical systems paper; it does not claim a first-principles derivation whose output is defined by its inputs. The closest thing to a self-referential loop is the latent critic QW, which is trained by distilling the action critic QA (Eq. 38) and then used as the actor's objective (Eq. 39). This is an explicitly disclosed training mechanism adopted from DSRL-NA, not a hidden fit masquerading as a prediction: the paper's reported outcomes are physical success, collision, and smoothness metrics (Tables 1 and 2) computed on held-out seen, unseen, and real-world rollouts with a shared MPC controller, not QW values. Reward normalization statistics (Eqs. 10–12) are fixed from the offline dataset and shared by all methods, which is ordinary preprocessing rather than a fitted-parameter prediction. The citation to M2Diffuser for the point-cloud representation is architectural and non-load-bearing; the comparison against M2Diffuser is empirical and M2Diffuser is externally published. No equation reduces by construction to another equation, and no fitted parameter is renamed as a predicted result. The absence of variance estimates or released code is a reproducibility/correctness concern, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

All declared assumptions and hand-chosen hyperparameters on which the central empirical claim rests. No new physical entities are introduced; the novel networks (latent steering network, latent critic) are architectural constructs rather than independent postulated entities.

free parameters (7)
  • Reward weights alpha_1..alpha_5 = (2.5, 2.0, 1.5, 1.5, 2.5)
    Hand-chosen to balance goal reaching, base limits, collision, smoothness, and joint limits; directly shapes what the RL critic learns (Eq. 13-14).
  • Flow-time logit-normal sampling distribution = N(-1.2, 1.2^2) via sigmoid
    Hand-chosen to emphasize intermediate noise states during flow pretraining (Eq. 34).
  • Action-critic warm-up steps N_pre = 20,000
    Hand-chosen; core stabilization contribution that postpones latent-critic and latent-actor updates.
  • Coarse-to-fine switch step t_cf = 50,000
    Hand-chosen; controls when the high-dimensional temporal residual branch is activated.
  • Residual scale lambda_res = 0.1
    Hand-chosen; scales the fine-grained latent residual when it is introduced.
  • Latent mixture coefficient rho = 0.5
    Hand-chosen; mixes the original flow prior with the steering distribution when training QW.
  • Reward shaping thresholds d_max, epsilon_c, kappa_max, clip c = 0.20 m, 0.05 m, 2.0 m^-1, 5
    Hand-chosen thresholds that define goal truncation, collision safety margin, base curvature limit, and reward clipping; these are closely related to the success criterion.
axioms (6)
  • standard math Rectified-flow regression loss learns a valid probability-flow ODE transporting Gaussian noise to the expert action-chunk distribution.
    Invoked without proof in the flow-policy pretraining section and Eq. 32-33, following Liu et al. 2023 and Lipman et al. 2023.
  • standard math The reparameterization trick provides unbiased gradients through sampled latent variables.
    Used in Algorithm 1 lines 18-19 and 29-31 to backpropagate through the Gaussian latent samples.
  • domain assumption Offline Q-learning with a target network and a fixed dataset gives a reasonable value estimate in this high-dimensional continuous setting.
    Standard RL assumption; no convergence guarantees are given for the 50x9 action-chunk space.
  • ad hoc to paper Value distillation from QA to QW preserves enough information for latent steering to improve decoded actions.
    The core mechanism of RLMM-Flow, assumed without formal analysis (Algorithm 1 lines 24-32, Eq. 38-39).
  • domain assumption Collision-sphere approximation plus ESDF distances captures real whole-body collision risk.
    Used throughout the reward formulation in the appendix; the approximation may miss thin obstacles or self-collision details.
  • domain assumption Robot-centric point-cloud representation is sufficient for sim-to-real transfer.
    Policies are trained on PhyScene point clouds and tested zero-shot on FAST-LIO2 reconstructed maps; the domain gap is assumed to be handled by the representation.

pith-pipeline@v1.3.0-daily-deepseek · 18477 in / 12724 out tokens · 117302 ms · 2026-08-01T15:20:50.847218+00:00 · methodology

0 comments
read the original abstract

Mobile manipulation requires generating whole-body action chunks that jointly satisfy goal reaching, collision avoidance, base kinematic constraints, manipulator joint limits, and trajectory smoothness. Flow-based generative policies provide an efficient paradigm for learning multimodal and temporally consistent motion priors from expert demonstrations, but imitation-only training cannot improve policy quality beyond the demonstration distribution. We propose RLMM-Flow, a flow-based mobile manipulation framework that combines expert flow-policy pretraining with latent-space reinforcement learning post-training. The framework first learns a flow policy that captures a multimodal whole-body motion prior from expert demonstrations. The pretrained flow policy is then frozen, while a latent steering network steers its initial noise toward higher-value action chunks. To stabilize high-dimensional latent optimization, we warm up an action-space critic before jointly training the latent critic and latent actor, and introduce coarse-to-fine latent steering that progressively expands control from a horizon-shared latent representation to a full-dimensional residual representation. Experiments on mobile manipulation motion-planning benchmarks show that RLMM-Flow substantially improves task success, collision avoidance, and trajectory quality over imitation-only flow policies and existing reinforcement learning post-training baselines, while preserving fast flow-based inference.

Figures

Figures reproduced from arXiv: 2607.26460 by Hui Cheng, Shuhang Wang, Ziming Li.

Figure 1
Figure 1. Figure 1: The Real Scene Performance of RLMM-Flow. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the RLMM-Flow training and infer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The complete high-dimensional temporal component [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of whole-body trajectories [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative simulation rollouts of RLMM-Flow [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evaluation reward curves for the QA warm-up and coarse-to-fine latent-steering ablations. maps are reconstructed using FAST-LIO2. All policies are transferred directly from simulation without real-world fine￾tuning and are executed with the same MPC controller. Fig￾ure 1 shows representative real-world rollouts. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_6.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

21 extracted references · 7 linked inside Pith

  1. [3]

    In Agrawal, P.; Kroemer, O.; and Burgard, W., eds., Proceedings of The 8th Conference on Robot Learning, vol- ume 270 ofProceedings of Machine Learning Research, 982–993

    Learning Robotic Manipulation Policies from Point Clouds with Conditional Flow Match- ing. In Agrawal, P.; Kroemer, O.; and Burgard, W., eds., Proceedings of The 8th Conference on Robot Learning, vol- ume 270 ofProceedings of Machine Learning Research, 982–993. PMLR. Garrett, C. R.; Chitnis, R.; Holladay, R.; Kim, B.; Silver, T.; Kaelbling,L.P.;andLozano-...

    2021

  2. [4]

    arXiv:2304.10573

    IDQL: Implicit Q-Learning as an Actor- Critic Method with Diffusion Policies. arXiv:2304.10573. Huang, S.; Wang, Z.; Li, P.; Jia, B.; Liu, T.; Zhu, Y.; Liang, W.; and Zhu, S.-C

    Pith/arXiv arXiv

  3. [5]

    In Agrawal, P.; Kroemer, O.; and Burgard, W., eds., Proceedings of The 8th Conference on Robot Learning, vol- ume 270 ofProceedings of Machine Learning Research, 1949–1974

    3D Diffuser Actor: Policy Diffusion with 3D Scene Representa- tions. In Agrawal, P.; Kroemer, O.; and Burgard, W., eds., Proceedings of The 8th Conference on Robot Learning, vol- ume 270 ofProceedings of Machine Learning Research, 1949–1974. PMLR. Khatib,O.1999.Mobilemanipulation:Theroboticassistant. Robotics and Autonomous Systems, 26(2-3): 175–183. Kost...

    1949

  4. [7]

    In The Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5,

    Flow Matching for Generative Modeling. In The Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  5. [8]

    Liu,X.;Gong,C.;andLiu,Q.2023

    OpenReview.net. Liu,X.;Gong,C.;andLiu,Q.2023. FlowStraightandFast: LearningtoGenerateandTransferDatawithRectifiedFlow. InThe Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  6. [9]

    McAllister, D.; Ge, S.; Yi, B.; Kim, C

    Motion planning around obstacles with convex optimization.Science robotics, 8(84): eadf7843. McAllister, D.; Ge, S.; Yi, B.; Kim, C. M.; Weber, E.; Choi, H.;Feng,H.;andKanazawa,A.2025. FlowMatchingPolicy Gradients. arXiv:2507.21053. Park, S.; Li, Q.; and Levine, S

    Pith/arXiv arXiv 2025

  7. [11]

    arXiv:2507.15073

    Reinforcement Learning for Flow-Matching Policies. arXiv:2507.15073. Psenka, M.; Escontrela, A.; Abbeel, P.; and Ma, Y

    Pith/arXiv arXiv

  8. [13]

    InInter- national Conference on Learning Representations, volume 2025, 77288–77329

    Diffusion policy policy optimization. InInter- national Conference on Learning Representations, volume 2025, 77288–77329. Schulman,J.;Ho,J.;Lee,A.X.;Awwal,I.;Bradlow,H.;and Abbeel, P

    2025

  9. [15]

    In2023 IEEE international conference on robotics and automation (ICRA), 5923–5930

    Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In2023 IEEE international conference on robotics and automation (ICRA), 5923–5930. IEEE. Wagenmaker, A.; Nakamoto, M.; Zhang, Y.; Park, S.; Yagoub,W.;Nagabandi,A.;Gupta,A.;andLevine,S.2025. SteeringYourDiffusionPolicywithLatentSpaceReinforce- ...

    Pith/arXiv arXiv 2025

  10. [16]

    InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    Diffusion Poli- cies as an Expressive Policy Class for Offline Reinforce- ment Learning. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  11. [18]

    Robotics, 38(4): 2053–2073

    FAST- LIO2: Fast Direct LiDAR-Inertial Odometry.IEEE Trans. Robotics, 38(4): 2053–2073. Yan, G.; Wu, Y.-H.; and Wang, X

    2053

  12. [19]

    arXiv:2305.13122

    Policy Representation viaDiffusionProbabilityModelforReinforcementLearning. arXiv:2305.13122. Yang, Y.; Jia, B.; Zhi, P.; and Huang, S

    Pith/arXiv arXiv

  13. [20]

    InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 16262–16272

    PhyScene: PhysicallyInteractable3DSceneSynthesisforEmbodiedAI. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 16262–16272. IEEE. Ze, Y.; Zhang, G.; Zhang, K.; Hu, C.; Wang, M.; and Xu, H

    2024

  14. [21]

    Zhang, Q.; Liu, Z.; Fan, H.; Liu, G.; Zeng, B.; and Liu, S. 2025a. FlowPolicy: Enabling Fast and Robust 3D Flow- BasedPolicyviaConsistencyFlowMatchingforRobotMa- nipulation. InWalsh,T.;Shah,J.;andKolter,Z.,eds.,Thirty- Ninth AAAI Conference on Artificial Intelligence, Thirty- Seventh Conference on Innovative Applications of Artificial Intelligence, Fiftee...

    2025

  15. [2009]

    In2009 IEEE International Confer- enceonRoboticsandAutomation,ICRA2009,Kobe,Japan, May 12-17, 2009, 489–494

    CHOMP: Gradient optimization techniques for effi- cient motion planning. In2009 IEEE International Confer- enceonRoboticsandAutomation,ICRA2009,Kobe,Japan, May 12-17, 2009, 489–494. IEEE. Ren, A.; Lidard, J.; Ankile, L.; Simeonov, A.; Agrawal, P.; Majumdar, A.; Burchfiel, B.; Dai, H.; and Simchowitz, M

    2009

  16. [2019]

    Pfrommer, S.; Huang, Y.; and Sojoudi, S

    Advantage-Weighted Regression: Simple and Scalable Off- Policy Reinforcement Learning.CoRR, abs/1910.00177. Pfrommer, S.; Huang, Y.; and Sojoudi, S

    Pith/arXiv arXiv 1910

  17. [2022]

    InThe Tenth International Conference on Learning Representa- tions, ICLR 2022, Virtual Event, April 25-29,

    Offline Re- inforcement Learning with Implicit Q-Learning. InThe Tenth International Conference on Learning Representa- tions, ICLR 2022, Virtual Event, April 25-29,

    2022

  18. [2023]

    In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 1916–1923

    Motion planning diffusion: Learning and planning of robot motions with diffusion models. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 1916–1923. IEEE. Chen, S.; Liu, J.; Qian, S.; Jiang, H.; Liu, Z.; Gu, C.; Li, X.; Hou, C.; Wang, P.; Wang, Z.; et al

    1916

  19. [2024]

    InIEEE International Con- ferenceonRoboticsandAutomation,ICRA2024,Yokohama, Japan, May 13-17, 2024, 1369–1375

    Real-time Whole-body Motion Planning for Mo- bile Manipulators Using Environment-adaptive Search and Spatial-temporal Optimization. InIEEE International Con- ferenceonRoboticsandAutomation,ICRA2024,Yokohama, Japan, May 13-17, 2024, 1369–1375. IEEE. Wu,X.;Lao,Y.;Jiang,L.;Liu,X.;andZhao,H.2022. Point Transformer V2: Grouped Vector Attention and Partition- b...

    2024

  20. [2025]

    Chi, C.; Xu, Z.; Feng, S.; Cousineau, E.; Du, Y.; Burchfiel, B.;Tedrake,R.;andSong,S.2025

    Ac-dit: Adap- tive coordination diffusion transformer for mobile manipu- lation.Advances in Neural Information Processing Systems, 38: 64008–64036. Chi, C.; Xu, Z.; Feng, S.; Cousineau, E.; Du, Y.; Burchfiel, B.;Tedrake,R.;andSong,S.2025. Diffusionpolicy:Visuo- motorpolicylearningviaactiondiffusion.TheInternational Journal of Robotics Research, 44(10-11):...

    2025

  21. [2026]

    arXiv:2606.01151

    Lagrangian Perturba- tion Diffusion Steering: Latent Reinforcement Learning for Generative Policies. arXiv:2606.01151. Urain, J.; Funk, N.; Peters, J.; and Chalvatzaki, G

    Pith/arXiv arXiv