Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Learning Visuotactile Estimation and Control for Non-prehensile Manipulation under Occlusions

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper claims a robot can keep pushing an object to target under prolonged visual occlusions by learning a visuotactile state estimator that reports its own uncertainty and training the control policy with that uncertainty in the loop.

desk verdict A credible systems paper: the components are known, but the explicit occlusion-focused formulation plus estimator-in-the-loop RL and positive hardware results make it a legitimate new result; the main soft spot is the unvalidated synthetic sensor model in sim-to-real. read the letter →

arxiv 2412.13157 v1 pith:TKRPVGMB submitted 2024-12-17 cs.RO cs.LG

classification cs.ROcs.LG
keywords non-prehensilemanipulationplanarpushingvisuotactilestateestimationBayesiandeeplearninguncertainty-awarecontrolreinforcementsim-to-realtransfervisualocclusions
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

The paper claims that a robot can keep pushing an object to a target even when its only camera view is occluded, by learning a state estimator that fuses vision, proprioception, and force/tactile readings and by feeding that estimator's uncertainty into the control policy. The estimator is trained on diverse simulated interaction data generated by privileged policy checkpoints, using a Bayesian deep learning loss that outputs both pose and covariance. The policy is then trained with the estimator in the loop, which lets it learn actions that actively reduce uncertainty, such as switching contact faces. On a planar pushing task the method reaches 94% success in simulation and transfers zero-shot to hardware, with 19/20, 10/10, and 7/10 runs under different occlusion scenarios. If this holds, non-prehensile manipulation no longer needs elaborate external tracking rigs, and a simple onboard camera plus force sensing may suffice.

What carries the argument

The machinery is a two-stage learning pipeline built on a GPU-accelerated physics simulator. A privileged policy $\pi_{\mathrm{priv}}(s_t)$ is trained with PPO in an occlusion-free environment, and equally spaced checkpoints (including untrained and suboptimal ones) are rolled out to collect 750,000 trajectories. Those trajectories are processed with a synthetic observation model that adds correlated and uncorrelated Gaussian noise and occlusions that freeze the last observed pose, with duration sampled from $\mathcal{N}(10, 5^2)$ seconds. The estimator is an LSTM-based network with MC dropout that predicts the object pose and aleatoric log-covariance; epistemic uncertainty is computed from 50 stochastic forward passes, and total uncertainty is $\hat{\Sigma}_t = \hat{\Sigma}^{\mathrm{ale}}_t + \hat{\Sigma}^{\mathrm{epi}}_t$. The control policy, an RL agent with the same recurrent architecture, receives the estimated state and the standard deviations of the total uncertainty as inputs, and is trained by PPO in the occlusion environment with the estimator frozen in the loop.

What would settle it

Observe whether the estimated uncertainty rises during gradual partial occlusions on hardware; a failure to anticipate full occlusion would indicate the binary freeze-last-pose observation model is load-bearing.

Watch

Extended reading notes

Core claim

The central discovery is that explicitly modeling both aleatoric and epistemic uncertainty in a learned visuotactile state estimator, and providing that uncertainty to a control policy trained with the estimator in the loop, makes planar pushing robust to prolonged visual occlusions. The estimator $f(o_t)$ outputs the object pose mean $\hat{q}^{\mathrm{obj}}_t$ and a diagonal covariance $\hat{\Sigma}_t$, trained with the negative log-likelihood loss $\mathcal{L} = \frac{1}{2}\ln|\hat{\Sigma}^{\mathrm{ale}}_t| + \frac{1}{2}(q^{\mathrm{obj}}_t - \hat{q}^{\mathrm{obj}}_t)^\top(\hat{\Sigma}^{\mathrm{ale}}_t)^{-1}(q^{\mathrm{obj}}_t - \hat{q}^{\mathrm{obj}}_t)$, with epistemic uncertainty added via Monte Carlo dropout. This estimator outperforms an MSE-trained variant (4.42 mm vs 7.62 mm mean translation error) and a vision-only baseline, and the policy $\pi_{\mathrm{est}}(\hat{s}_t, \hat{\Sigma}_t)$ reaches 94% success versus 12% for end-to-end RL and 38% for behavior cloning. The analysis of contact behavior shows the policy learns to switch contact faces to reduce estimator uncertainty, something the privileged policy applied to estimated states does less effectively (83% success).

Load-bearing premise

The load-bearing premise is that the synthetic observation model used in training—occlusions that freeze the last observed pose, Gaussian noise levels in Table 1, and neglect of pusher dynamics compensated by large synthetic noise—represents the real robot's sensors and occlusion events closely enough that zero-shot sim-to-real transfer works.

Editorial extensions

If this is right

  • Onboard-only perception becomes sufficient for planar non-prehensile manipulation under occlusion, removing the need for Vicon, under-table cameras, or multi-camera rigs.
  • Uncertainty-aware policies naturally learn information-gathering behaviors (e.g., switching contact faces) that reduce state uncertainty, improving task success beyond what nominal estimates provide.
  • Explicit state estimation with uncertainty beats both end-to-end policy learning and teacher-student behavior cloning for prolonged occlusion scenarios in this task.
  • Zero-shot sim-to-real transfer is possible even when the simulator omits pusher dynamics, provided observation noise is randomized aggressively during training.

Reading between the lines

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

  • The uncertainty estimates could be repurposed for failure prediction and fallback triggering, and for active data collection to improve estimator coverage; the authors list these as future work but the machinery already produces the quantities needed.
  • The synthetic occlusion model—freeze the last pose—is likely the tightest bottleneck; testing with partial or gradual occlusions on hardware would reveal whether the learned uncertainty calibration generalizes beyond that specific model.
  • The dependency on a fixed cuboid object and spherical pusher suggests the method's claims about uncertainty handling are demonstrated for one geometry; extending to diverse shapes would require re-collection of privileged data or a shape-conditioned estimator.
  • A direct comparison with a model-based filter (e.g., a particle filter fusing force and vision) on the same occlusion scenarios would clarify whether the learned uncertainty adds value beyond classical estimation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a visuotactile state estimation and uncertainty-aware control pipeline for planar non-prehensile manipulation under visual occlusions. The method first trains a privileged RL policy in an occlusion-free simulation, uses its training checkpoints to collect a diverse dataset, and then trains a Bayesian deep learning estimator that outputs both object pose and aleatoric/epistemic uncertainty via MC dropout. The estimator is embedded in the RL loop to train a control policy that receives the estimated pose and covariance as input. In simulation the method achieves 94% success, and with zero-shot sim-to-real transfer it attains 19/20, 10/10, and 7/10 success on a KUKA iiwa with an onboard camera under naturally occurring and human-induced occlusions. The central claim is that explicitly modeling uncertainty, rather than relying on complex external perception, is what makes occlusion-robust behavior possible after transfer.

Significance. If the result holds, the paper makes a valuable contribution: it demonstrates that learned uncertainty from a proprioceptive/tactile estimator can substitute for continuous visual tracking in a contact-rich non-prehensile task, enabling simpler onboard perception setups. The experimental design is generally careful: separate train/validation/test datasets, three RL training seeds, multiple ablations (MSE estimator, likelihood loss, end-to-end RL, behavior cloning, privileged policy with estimator), and hardware trials with randomized initial configurations. The paper also provides reproducible training details in the appendix. The main open question is whether the hardware success is actually attributable to the learned uncertainty mechanism, given that the uncertainty calibration is not evaluated on the robot and the simulation observation model is admittedly approximate.

major comments (4)
  1. [Section 6 and Section 7] The hardware section states that the simulator neglects pusher dynamics that cause 'significant force readings outside of contact interactions' on the real F/T sensor and that the authors instead 'relied on adding large amounts of correlated and uncorrelated noise to the force observation.' Section 7 additionally states that 'we omit any quantitative evaluation on the accuracy of the estimated uncertainty.' Because the policy receives the predicted covariance from Eq. (4) as an input (Section 3.3) and the authors attribute the successful behavior to uncertainty-aware contact-face switching (Section 5.4), the hardware results do not currently establish that the learned uncertainty is well-calibrated on the real robot. Without measuring estimator pose error or covariance calibration against ground truth on the hardware, the positive trials only show that the overall system works, not that the synthetic occlusion/noise model is representative enough for the claimed mechanism to be the cause of success. Please add a quantitative hardware evaluation of the estimator's accuracy and uncertainty calibration (e.g., comparing predicted variance with empirical error across trials) or substantively temper the claim that uncertainty modeling is what enables the sim-to-real transfer.
  2. [Table 2] The claim that modeling uncertainty 'significantly improves the state estimator accuracy' is based solely on Table 2, which reports point estimates (Mean L2, RMSE, Mean Abs., RMSE) with no variance, no number of seeds, and no statistical test. Given that the estimator is the foundation of the policy and that the differences between MSE (7.62 mm), Likelihood (6.83 mm), and Ours (4.42 mm) are the basis for this claim, please report standard deviations or confidence intervals across estimator training seeds and test splits, or provide a significance test.
  3. [Section 5.2 and Figure 2] The distinction between πest(ŝ, Σ) and πest(ŝ) supports the paper's emphasis on explicit uncertainty input, but the final difference is only 94% versus 92% and Figure 2 shows overlapping standard deviations across three seeds. The conclusion that 'providing the uncertainty estimate explicitly to the policy leads to slightly improved performance' is not supported by any significance test. Please report per-seed final success rates and perform a simple statistical comparison (e.g., bootstrap or a paired test across seeds), or soften the claim to a trend.
  4. [Section 3.2 Data Processing and Section 6] The synthetic occlusion model freezes the last observed object pose for a duration sampled from N(10, 5^2) s and adds a binary indicator ξ. On the real hardware, occlusions arise when AprilTag markers leave the camera field of view or are obstructed by the pusher. The paper does not specify how the real detection pipeline behaves during occlusion (does it output the last detected pose, a missing flag, or a zero vector?) and whether that behavior matches the synthetic 'frozen pose' model. This is a potential sim-to-real gap in the observation model that affects the estimator's inputs and the policy's behavior. Please clarify the actual detection pipeline and, ideally, report the empirical distribution of detection outages on the hardware to justify the synthetic model.
minor comments (6)
  1. [Table 1] Table 1 lists the occlusion duration as N(10, 5^2) s, but Section 5.3 fixes the duration for evaluation scenarios. Please clarify whether the table refers only to training, and separately state the distribution of the correlated noise component referenced in Section 4.
  2. [Eq. (2)] Equation (2) gives the sample covariance without Bessel's correction (the outer factor is 1/M rather than 1/(M−1)). For M=50 this is numerically negligible, but the formula should be flagged as the biased estimator or corrected for formal consistency.
  3. [Section 4 (State Estimator)] The paper says '300 uniformly spaced privileged policy checkpoints' are used for data collection, but it is not stated how spacing is measured (e.g., every N optimizer steps) or whether the checkpoints include the final trained policy. Please specify the selection criterion.
  4. [Section 6] The hardware success counts 19/20, 10/10, and 7/10 are small samples; please report binomial (e.g., Clopper-Pearson) confidence intervals to convey the precision of these rates.
  5. [Section 5.1] The comparison with prior model-based estimators [36, 37] is qualitative ('similar or better accuracy'); please provide the relevant numeric accuracies from those references so the reader can verify the claim.
  6. [Section 5.4] In the sentence 'We find that πpriv(st) makes 2.68±1.53, while πest(ŝt, Σt) makes 4.22±2.67 contact switches per episode,' the notation mixes πpriv(st) and πpriv(ŝt); please use consistent notation for the privileged policy with and without the estimator.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the estimator and policy are trained and evaluated on independent ground-truth data, and self-citations only support implementation choices rather than load-bearing claims.

full rationale

The paper's derivation chain is self-contained rather than circular. The state estimator is trained from privileged-policy trajectories to which synthetic occlusions and noise are applied, and its accuracy is evaluated on a separate held-out test dataset using ground-truth object poses, not the estimator's own output. The control policy is trained with the pre-trained estimator in the loop, but the reward and the reported success metrics are computed from the environment's ground-truth state, so neither the estimator nor the policy is used to fit any reported constant. The hardware experiments provide an external benchmark: the trained estimator and policy are deployed zero-shot on the physical robot with real occlusions, so the central claim does not reduce to a re-evaluation of training data. The paper cites prior work by the same authors, notably [11] for categorical action exploration and [38] for the optimization-based task specification library, but these citations support implementation details and are not load-bearing premises or uniqueness arguments. The acknowledged limitation that uncertainty accuracy was not quantitatively evaluated on hardware is an empirical validity concern, not a circularity concern: it does not mean any prediction was defined in terms of its input or that a fitted parameter was renamed as a prediction. No specific circular step can be quoted or exhibited from the paper's equations or claims.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claims depend on a small set of hand-chosen distributions for occlusions, noise, and dynamics randomization, plus standard Bayesian deep learning approximations. No constants are fitted to the target success metric, so the circularity burden is low. The evaluation regime is largely defined by these parameters, and the hardware results are the only external check.

free parameters (6)
  • Occlusion start probability p = 1/30 per timestep
    Chosen by hand to create frequent occlusions; defines the difficulty of the trained and tested task (Section 3.2, Table 1).
  • Occlusion duration distribution = N(10 s, 5^2 s^2)
    Synthetic occlusion lengths in training and evaluation; the reported success rates depend on this regime (Table 1).
  • Observation noise standard deviations = force 0.7 N, position 0.0025 m, orientation 0.05 rad
    Added to synthetic observations; the estimator and policy are trained and evaluated under these noise levels (Table 1).
  • Dynamics randomization ranges = mass U(3,3.5) kg, static friction U(0.3,0.5), dynamic friction U(0.1,0.3), restitution U(0.1,0.7)
    Chosen to support sim-to-real; defines the domain variation the method is exposed to (Table 3).
  • Reward weights k1, k2, k3 = 0.1, 0.02, 0.004
    Shape the task reward; success is scored by distance thresholds (1 cm, 15 deg) rather than these weights, but policy learning is affected (Section 4).
  • Monte Carlo dropout samples M = 50
    Number of stochastic forward passes used for epistemic uncertainty and predictive mean (Section 3.2).
assumptions (6)
  • standard math MC dropout approximates Bayesian variational inference
    Used to estimate epistemic uncertainty and the predictive mean (Section 3.2, Eqs. 2-4; cited to Gal and Ghahramani 2016).
  • standard math Aleatoric and epistemic uncertainties are independent and additively combined
    Total covariance is the sum of aleatoric and epistemic terms (Eq. 4), following Kendall and Gal 2017; an approximation, not proven for this setting.
  • domain assumption Object pose occlusions freeze at the last observed value and a binary occlusion indicator is available
    The synthetic observation model used for training and evaluation; real occlusion behavior may differ (Section 3.2).
  • domain assumption End-effector pose and force measurements are always available and unoccluded
    Assumed in the problem formulation; on hardware, unmodeled pusher dynamics corrupt force readings, acknowledged in Section 6.
  • domain assumption Planar pushing with a fixed cuboid object and spherical pusher captures the key challenges of non-prehensile manipulation
    Task and geometry are fixed; generalization to other shapes or tasks is not demonstrated (Section 4, Limitations).
  • domain assumption Dynamics randomization ranges in Table 3 are sufficient for zero-shot sim-to-real transfer
    No systematic hardware uncertainty calibration is provided; supported only by task success counts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Visuotactile Estimation and Control for Non-prehensile Manipulation under Occlusions." pith.science (2026). https://pith.science/paper/TKRPVGMB

@misc{pith2026241213157,
  author       = {Pith},
  title        = {Pith review of: Learning Visuotactile Estimation and Control for Non-prehensile Manipulation under Occlusions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKRPVGMB}},
  note         = {Machine review of arXiv:2412.13157}
}
read the original abstract

Manipulation without grasping, known as non-prehensile manipulation, is essential for dexterous robots in contact-rich environments, but presents many challenges relating with underactuation, hybrid-dynamics, and frictional uncertainty. Additionally, object occlusions in a scenario of contact uncertainty and where the motion of the object evolves independently from the robot becomes a critical problem, which previous literature fails to address. We present a method for learning visuotactile state estimators and uncertainty-aware control policies for non-prehensile manipulation under occlusions, by leveraging diverse interaction data from privileged policies trained in simulation. We formulate the estimator within a Bayesian deep learning framework, to model its uncertainty, and then train uncertainty-aware control policies by incorporating the pre-learned estimator into the reinforcement learning (RL) loop, both of which lead to significantly improved estimator and policy performance. Therefore, unlike prior non-prehensile research that relies on complex external perception set-ups, our method successfully handles occlusions after sim-to-real transfer to robotic hardware with a simple onboard camera. See our video: https://youtu.be/hW-C8i_HWgs.

Figures

Figures reproduced from arXiv: 2412.13157 by the authors.

Figure 1
Figure 1. Snapshots from an exemplar robot motion pushing a box to the target with occlusions. Columns [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Performance with different training configurations [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Performance of different control policies with in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Trajectories generated under full occlusion. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Planar pushing simulation environment in Isaac Sim. The pusher is [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Training performance of the privileged policy [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Estimator validation loss when training with the [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Estimator validation loss when training with the [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Behavior cloning policy validation loss. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 20 canonical work pages

  1. [1]

    M. T. Mason. Mechanics and planning of manipulator pushing operations. The International Journal of Robotics Research, 5(3):53–71, 1986. doi:10.1177/027836498600500303

  2. [2]

    F. R. Hogan and A. Rodriguez. Reactive planar non-prehensile manipulation with hybrid model predictive control. The International Journal of Robotics Research, 39(7):755–773, 2020. doi: 10.1177/0278364920913938

  3. [3]

    Moura, T

    J. Moura, T. Stouraitis, and S. Vijayakumar. Non-prehensile planar manipulation via trajectory optimization with complementarity constraints. In 2022 IEEE International Conference on Robotics and Automation (ICRA), pages 970–976. IEEE, 2022. doi:10.1109/ICRA46639.2022. 9811942

  4. [4]

    J. Zhou, R. Paolini, A. M. Johnson, J. A. Bagnell, and M. T. Mason. A probabilistic planning framework for planar grasping under uncertainty. IEEE Robotics and Automation Letters , 2 (4):2111–2118, 2017. doi:10.1109/LRA.2017.2720845

  5. [5]

    Bauza and A

    M. Bauza and A. Rodriguez. A probabilistic data-driven model for planar pushing. In 2017 IEEE International Conference on Robotics and Automation (ICRA) , pages 3008–3015. IEEE,

  6. [6]

    W. Zhou, B. Jiang, F. Yang, C. Paxton, and D. Held. Hacman: Learning hybrid actor-critic maps for 6d non-prehensile manipulation. In Conference on Robot Learning , pages 241–265. PMLR, 2023

  7. [7]

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3803–3810. IEEE, 2018. doi:10.1109/ICRA.2018.8460528

  8. [8]

    Lowrey, S

    K. Lowrey, S. Kolev, J. Dao, A. Rajeswaran, and E. Todorov. Reinforcement learning for non- prehensile manipulation: Transfer from simulation to physical system. In IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR) , pages 35–42, 2018. doi:10.1109/SIMPAR.2018.8376268

Show all 40 references
  1. [9]

    Jeong, J

    R. Jeong, J. Kay, F. Romano, T. Lampe, T. Rothorl, A. Abdolmaleki, T. Erez, Y . Tassa, and F. Nori. Modelling generalized forces with reinforcement learning for sim-to-real transfer. arXiv preprint arXiv:1910.09471, 2019

  2. [10]

    L. Cong, H. Liang, P. Ruppel, Y . Shi, M. G ¨orner, N. Hendrich, and J. Zhang. Reinforce- ment learning with vision-proprioception model for robot planar pushing. Frontiers in Neuro- robotics, 16, 2022. ISSN 1662-5218. doi:10.3389/fnbot.2022.829437

  3. [11]

    Del Aguila Ferrandis, J

    J. Del Aguila Ferrandis, J. Moura, and S. Vijayakumar. Nonprehensile planar manipulation through reinforcement learning with multimodal categorical exploration. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 5606–5613. IEEE,

  4. [12]

    Zhou and D

    W. Zhou and D. Held. Learning to grasp the ungraspable with emergent extrinsic dexterity. In Conference on Robot Learning, pages 150–160. PMLR, 2023

  5. [13]

    S.-M. Yang, M. Magnusson, J. A. Stork, and T. Stoyano. Learning extrinsic dexterity with pa- rameterized manipulation primitives. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5404–5410. IEEE, 2024. doi:10.1109/ICRA57147.2024.10611431. 9

  6. [15]

    G. Liu, J. De Winter, D. Steckelmacher, R. K. Hota, A. Nowe, and B. Vanderborght. Synergistic task and motion planning with reinforcement learning-based non-prehensile actions. IEEE Robotics and Automation Letters, 2023. doi:10.1109/LRA.2023.3261708

  7. [17]

    Y . Cho, J. Han, Y . Cho, and B. Kim. CORN: Contact-based Object Representation for Non- prehensile Manipulation of General Unseen Objects. In International Conference on Learning Representations (ICLR), 2024

  8. [18]

    Wang and E

    J. Wang and E. Olson. Apriltag 2: Efficient and robust fiducial detection. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 4193–4198. IEEE,

  9. [19]

    Patil, Y

    S. Patil, Y . Duan, J. Schulman, K. Goldberg, and P. Abbeel. Gaussian belief space planning with discontinuities in sensing domains. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6483–6490. IEEE, 2014. doi:10.1109/ICRA.2014.6907816

  10. [20]

    G. Chou, N. Ozay, and D. Berenson. Safe output feedback motion planning from im- ages via learned perception modules and contraction theory. In International Workshop on the Algorithmic F oundations of Robotics , pages 349–367. Springer, 2022. doi:10.1007/ 978-3-031-21090-7 21

  11. [21]

    Andrychowicz, B

    OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, et al. Learning dexterous in-hand manipu- lation. The International Journal of Robotics Research , 39(1):3–20, 2020. doi:10.1177/ 0278364919887447

  12. [22]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. InConference on Robot Learning, pages 91–100. PMLR, 2022

  13. [23]

    Calandra, A

    R. Calandra, A. Owens, D. Jayaraman, J. Lin, W. Yuan, J. Malik, E. H. Adelson, and S. Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. IEEE Robotics and Automation Letters, 3(4):3300–3307, 2018. doi:10.1109/LRA.2018.2852779

  14. [24]

    Zhang, M

    K. Zhang, M. Sharma, M. Veloso, and O. Kroemer. Leveraging multimodal haptic sensory data for robust cutting. In2019 IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids), pages 409–416. IEEE, 2019. doi:10.1109/Humanoids43949.2019.9035073

  15. [25]

    Y . Hao, R. Wang, Z. Cao, Z. Wang, Y . Cui, and D. Sadigh. Masked imitation learning: Dis- covering environment-invariant modalities in multimodal demonstrations. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 1–7. IEEE, 2023. doi:10...

  16. [26]

    M. A. Lee, Y . Zhu, K. Srinivasan, P. Shah, S. Savarese, L. Fei-Fei, A. Garg, and J. Bohg. Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. In 2019 IEEE International Conference on Robotics and Automation (ICRA), ...

  17. [27]

    D. Chen, B. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl. Learning by cheating. In Conference on Robot Learning, pages 66–75. PMLR, 2020. 10

  18. [28]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning robust percep- tive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62):eabk2822, 2022. doi:10.1126/scirobotics.abk2822

  19. [29]

    T. Chen, J. Xu, and P. Agrawal. A system for general in-hand object re-orientation. In Confer- ence on Robot Learning , pages 297–307. PMLR, 2022

  20. [30]

    Kendall and Y

    A. Kendall and Y . Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems , 30, 2017

  21. [31]

    R. L. Russell and C. Reale. Multivariate uncertainty in deep learning. IEEE Transactions on Neural Networks and Learning Systems, 33(12):7937–7943, 2021. doi:10.1109/TNNLS.2021. 3086757

  22. [32]

    Gal and Z

    Y . Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncer- tainty in deep learning. In International Conference on Machine Learning , pages 1050–1059. PMLR, 2016

  23. [33]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research , 15 (1):1929–1958, 2014

  24. [34]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021

  25. [35]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

  26. [36]

    Yu and A

    K.-T. Yu and A. Rodriguez. Realtime state estimation with tactile and visual sensing. ap- plication to planar manipulation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 7778–7785. IEEE, 2018. doi:10.1109/ICRA.2018.8463183

  27. [37]

    M. A. Lee, B. Yi, R. Mart ´ın-Mart´ın, S. Savarese, and J. Bohg. Multimodal sensor fusion with differentiable filters. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10444–10451. IEEE, 2020. doi:10.1109/IROS45743.2020.9341579

  28. [38]

    C. E. Mower, J. Moura, N. Z. Behabadi, S. Vijayakumar, T. Vercauteren, and C. Bergeles. Op- tas: An optimization-based task specification library for trajectory optimization and model pre- dictive control. In 2023 IEEE International Conference on Robotics and Automation (ICRA)...

  29. [39]

    Heess, D

    N. Heess, D. Tb, S. Sriram, J. Lemmon, J. Merel, G. Wayne, Y . Tassa, T. Erez, Z. Wang, S. Eslami, et al. Emergence of locomotion behaviours in rich environments. arXiv preprint arXiv:1707.02286, 2017

  30. [40]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 11 Appendix A Simulation Environment We develop our planar pushing simulation environment using NVIDIA Omniverse Isaac Sim due to its GPU parallelization capabilities, whi...

  31. [2016]

    doi:10.1109/IROS.2016.7759617

  32. [2017]

    doi:10.1109/ICRA.2017.7989345

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.