Pith. sign in

REVIEW 2 major objections 6 minor 28 references

This paper proposes Bench-Push, the first unified benchmark for pushing-based mobile robot navigation and manipulation, with new metrics for efficiency, effort, and partial success.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 16:46 UTC pith:W4BLMPNL

load-bearing objection A genuinely useful, well-scoped benchmark for pushing-based navigation and manipulation, with one headline metric (E_manip) that needs rethinking but not fatal; worth a serious referee. the 2 major comments →

arxiv 2512.11736 v2 pith:W4BLMPNL submitted 2025-12-12 cs.RO

Bench-Push: Benchmarking Pushing-based Navigation and Manipulation Tasks for Mobile Robots

Ninghan Zhong , Steven Caro , Megnath Ramesh , Rishi Bhatnagar , Avraiem Iskandar , Stephen L. Smith This is my paper
classification cs.RO
keywords benchmarkpushing-based navigationpushing-based manipulationevaluation metricssim-to-real transfermovable obstaclesnon-prehensile pushingpartial task completion
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 argues that pushing-based mobile robot research has been held back by ad hoc environments and inconsistent metrics, and that a shared benchmark can fix that. It introduces Bench-Push, which covers four task families—maze navigation with movable obstacles, ship navigation through ice, box delivery, and area clearing—each in fast 2D and realistic 3D simulation. The benchmark's core contribution is a set of metrics that separate how efficiently a robot reaches a goal or completes a manipulation from how much pushing effort it expends, and that credit partial task completion. The paper also provides reference policies and reports that policies trained in Bench-Push's 3D simulation transfer to a physical robot with consistent ranking of algorithms.

Core claim

The central claim is that a single benchmark can support both navigation-centric and manipulation-centric pushing tasks while measuring the qualities that matter for pushing: path efficiency, interaction effort, and partial completion. Bench-Push implements this with five metrics: for navigation, E_nav compares the robot's path length to the shortest collision-free path while I_nav is the ratio of work spent moving the robot to work spent moving all objects; for manipulation, S_manip is the fraction of objects successfully delivered or cleared, E_manip compares the executed path to an idealized minimum path, and I_manip is the ratio of necessary work to total work. By providing baselines and

What carries the argument

The carrying mechanism is the benchmark's metric design plus its environment abstraction. The navigation metrics use work ratios derived from masses, path lengths, and a constant friction coefficient; the manipulation metrics add a success fraction and a lower-bound path length L*(O') computed from a minimum spanning tree over a graph of object starts, goal contact points, and the robot. This MST estimate is what allows E_manip to approximate 'idealized' efficiency for a partially completed multi-object task, though the paper notes it is not a strict lower bound. Around this, the benchmark provides a standard reinforcement-learning interface, interchangeable bumpers, wheeled and unwheeled bo

Load-bearing premise

The manipulation efficiency score E_manip relies on a minimum-spanning-tree estimate of the ideal path length that the paper admits is not a strict lower bound, so if that estimate is not reliably tight, E_manip may no longer mean what it claims.

What would settle it

Build or simulate a box-delivery instance where the true optimal pushing path for the completed sub-tasks is shorter than the MST estimate; if E_manip then exceeds 1 or the estimate fails to lower-bound the actual path, the metric's interpretation breaks.

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

If this is right

  • Researchers can compare pushing policies on identical tasks and score them on efficiency–effort trade-offs instead of success alone.
  • The 2D simulation tier lets users prototype quickly before committing to expensive 3D or hardware experiments.
  • Because interaction effort is scored separately from efficiency, policies that nudge obstacles lightly can be distinguished from those that bulldoze through them.
  • Bench-Push's reported sim-to-real consistency in Maze and Box-Delivery suggests the benchmark can serve as a training-and-evaluation loop for real deployment.

Where Pith is reading between the lines

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

  • A natural extension is to certify the L*(O') lower bound—for small instances one could solve exact shortest-push sequences and place a hard bound under E_manip; if the MST estimate can exceed 1 or fail to lower-bound, the metric would need renormalization.
  • The work-ratio framing could generalize to other non-prehensile contact tasks, such as sweeping, towing, or excavating, where efficiency of the robot's own motion versus effort applied to the environment is the natural comparison.
  • The current bird's-eye, known-pushable-object assumption implies the benchmark suits structured lab scenarios; adding egocentric perception and pushability estimation would extend it to partially known or cluttered real-world settings.

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

2 major / 6 minor

Summary. The paper introduces Bench-Push, a benchmark for pushing-based mobile robot navigation and manipulation. It provides four configurable environments in both 2D (Pymunk) and 3D (MuJoCo) simulation—Maze, Ship-Ice, Box-Delivery, and Area-Clearing—together with a set of novel evaluation metrics (E_nav, I_nav, S_manip, E_manip, I_manip). The benchmark includes reference baselines (SAC, PPO, RRT, SAM, ASV planners) and reports evaluations in simulation and on a physical TurtleBot testbed, claiming zero-shot sim-to-real transfer. The central claim is that this is the first unified benchmark for pushing-based mobile robot navigation and manipulation tasks.

Significance. If the metrics are sound, Bench-Push fills a recognized gap: the field of pushing-based robotics lacks standardized environments and evaluation protocols. The 2D/3D dual-simulation design, configurable environment variations, and bundled baseline implementations are practical contributions that could improve reproducibility and cross-comparison. The open-source Python library and the physical testbed demonstration are useful resources. The paper is transparent about limitations, and the environment descriptions are detailed enough for reimplementation. However, the significance of the contribution depends on the validity of the proposed metrics, particularly E_manip, whose reference path is computed with a heuristic that the authors admit is not a strict lower bound. The sim-to-real claim is also supported by limited data. These issues need to be addressed before the benchmark can be relied upon as a standard evaluation tool.

major comments (2)
  1. [Sec. III-C.2, Eq. (4)] The manipulation efficiency metric E_manip is defined as L*(O')/l0, where L*(O') is computed as the weight of an MST on a graph whose edges are restricted to object-object, robot-object, and object-own-goal pairs. As the authors state, this MST is not a strict lower bound on the true optimal path length, so E_manip can exceed 1 and its bias can vary with the object layout. This undermines the metric's comparability across episodes and its interpretation as an 'efficiency' score. Since novel metrics are a headline contribution, this is load-bearing. The authors should either replace L* with a true lower bound (e.g., solve an appropriate TSP/Steiner-tree problem on a graph with all relevant transitions) or redefine E_manip as a structural score with an established range, validating its behavior across random layouts. At minimum, an empirical study of E_manip's distribution over the benchma
  2. [Sec. IV-C/IV-D, Tables III-IV] The claim that 'policies trained in Bench-Push transfer well to real robots' is supported by only 3 physical episodes per condition, with no statistical tests. The reported rankings are not always consistent between simulation and testbed: in Table III, 10-Obs, RRT has the highest E_nav in simulation (0.758) but the lowest in testbed (0.681), whereas PPO is second in simulation and first in testbed. In Table IV, 5-Obs, SAM's E_manip is higher in the testbed (0.417) than in simulation (0.243), which is not a clear 'alignment'. The conclusion of consistent sim-to-real behavior is stronger than the evidence. Please temper the claim to a feasibility demonstration, or add more trials and appropriate statistical analysis (e.g., hypothesis tests or confidence intervals on rankings).
minor comments (6)
  1. [Sec. III-C.2] The phrase '(a lower bound on)' in the definition of L*(O') is misleading given the subsequent admission that the MST is not a strict lower bound. Clarify the wording.
  2. [Sec. IV-B] The physical testbed uses an overhead camera, but the policies expect egocentric observations (occupancy, distance transforms, heading encoding). The paper does not describe how the overhead detections are converted into these observations. This is necessary for reproducibility of the sim-to-real experiments.
  3. [Tables III and IV] The labels '3-Obs', '6-Obs', '10-Obs' are not defined in the captions or text. State explicitly that these denote the number of movable obstacles/boxes.
  4. [Sec. III-C.1, Eq. (2)] The denominator in Eq. (2) uses 'i=0' to include the robot, but the preceding text defines m_i and l_i only for objects i=1,...,K. Introduce m_0 and l_0 explicitly before the equation.
  5. [Sec. I] The abstract says the library 'will be open-sourced', while the body (Sec. I) says 'will also be made available' and the repository URL is given. Use consistent tense and clarify the current availability status of the code.
  6. [References] There are minor formatting inconsistencies in the bibliography, e.g., 'V .' instead of 'V.', and some entries lack page ranges or venue formatting. A thorough proofread of the references is advisable.

Circularity Check

0 steps flagged

No significant circularity: Bench-Push is a self-contained benchmark with metric definitions and external baselines.

full rationale

The paper's central contributions are new environments, metrics, and baseline evaluations. The metrics E_nav, I_nav, S_manip, E_manip, and I_manip are defined directly from path lengths, masses, and task-completion indicators, with no parameters fitted to the data being evaluated. The E_manip reference L*(O') is admittedly computed via an MST that is not a strict lower bound, but the paper explicitly acknowledges this limitation rather than disguising it as a proven quantity; this is a correctness/validity concern, not circularity. The Ship-Ice environment draws parameters from the authors' prior ASV work [14], [15], and their planners are used as task-specific baselines, but the benchmark's claims do not reduce to those citations: the baselines are independently implemented and evaluated, and the sim-to-real results compare trained policies against physical testbed outcomes. There is no fitted quantity being renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in through self-citation. The derivation chain is open and externally checkable, so the appropriate finding is no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No numerical parameters are fitted to data in this paper: l*_0 and l*_i are geometric quantities, and masses/drag coefficients are taken from cited prior work or set as simulation configuration. The load-bearing assumptions are about simulator fidelity, the friction-work model for the metrics, the ad-hoc MST bound, and the representativeness of the four chosen environments. No new physical entities are introduced.

axioms (5)
  • domain assumption Rigid-body contact dynamics in Pymunk and MuJoCo approximate real robot-object pushing sufficiently for policy transfer.
    Central to the sim-to-real claim; only tested on Maze and Box-Delivery under overhead-camera perception, not on onboard sensing or Ship-Ice.
  • domain assumption Ship-Ice fluid effects can be approximated by combined linear and quadratic drag on bodies in the xy-plane.
    Stated in Sec III-A.2; no validation of the drag approximation against real ship-ice interaction is provided.
  • domain assumption The work to move an object i a distance l_i is μ m_i g l_i with a single constant coefficient of kinetic friction μ across the environment.
    Assumed in Sec III-C for I_nav and I_manip; μ cancels in the ratios, but proportionality of effort to mass times path length is still assumed and object rotations and actuation losses are ignored.
  • ad hoc to paper The MST of the constructed graph is a practical lower bound on the minimum path length to complete the successful sub-tasks.
    The paper itself states the MST is not a strict lower bound and E_manip is not guaranteed to lie in [0,1], so this assumption is load-bearing for the manipulation efficiency metric.
  • ad hoc to paper The four selected environments (Maze, Ship-Ice, Box-Delivery, Area-Clearing) capture the fundamental challenges in pushing-based mobile robot navigation and manipulation.
    The representativeness of the benchmark is asserted in Sec I and III-A without a formal survey or coverage metric; it rests on the authors' selection from existing literature.

pith-pipeline@v1.3.0-alltime-deepseek · 11824 in / 13818 out tokens · 132061 ms · 2026-08-03T16:46:49.371713+00:00 · methodology

0 comments
read the original abstract

Mobile robots are increasingly deployed in cluttered environments with movable objects, posing challenges for traditional methods that prohibit interaction. In such settings, the mobile robot must go beyond traditional obstacle avoidance, leveraging pushing or nudging strategies to accomplish its goals. While research in pushing-based robotics is growing, evaluations rely on ad hoc setups, limiting reproducibility and cross-comparison. To address this, we present Bench-Push, the first unified benchmark for pushing-based mobile robot navigation and manipulation tasks. Bench-Push includes multiple components: 1) a comprehensive range of simulated environments that capture the fundamental challenges in pushing-based tasks, including navigating a maze with movable obstacles, autonomous ship navigation in ice-covered waters, box delivery, and area clearing, each with varying levels of complexity; 2) novel evaluation metrics to capture efficiency, interaction effort, and partial task completion; and 3) demonstrations using Bench-Push to evaluate example implementations of established baselines across environments. Bench-Push is open-sourced as a Python library with a modular design. The code, documentation, and trained models can be found at https://github.com/IvanIZ/BenchNPIN.

Figures

Figures reproduced from arXiv: 2512.11736 by Avraiem Iskandar, Megnath Ramesh, Ninghan Zhong, Rishi Bhatnagar, Stephen L. Smith, Steven Caro.

Figure 1
Figure 1. Figure 1: Maze environment in 2D simulation (top-left panel), in 3D simulation (lower-left panel), and in our physical testbed (right panel). and manipulation, where the mobile robot must interact with objects through pushing or nudging actions to progress. A parallel branch of work has traditionally studied such scenarios by assuming that mobile robots are equipped with manipulators and can use grasping actions to … view at source ↗
Figure 2
Figure 2. Figure 2: Illustrations of pushing-based navigation and manipulation environments in both 2D and 3D simulations, each with configurable [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The Maze environment (left) and an example egocentric observation for Maze (right). spaces and walls [17], [6], [4]. These environments capture common structural layouts of many real-world indoor set￾tings, such as offices and hospitals. Motivated by this, we introduce Maze environment, which features a static maze structure with randomly initialized obstacles ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The Box-Delivery environment (left) and an example obser￾vation for Box-Delivery (right). Surface Vehicle (ASV) navigation through ice-covered wa￾ters [14], [15]. This domain presents challenges distinct from Maze, such as irregular obstacles, dense clustering, and fluid dynamics. For this reason, we introduce the Ship-Ice environment, which simulates an autonomous ship navigating in a channel of ice-cover… view at source ↗
Figure 5
Figure 5. Figure 5: Three teleoperated Maze paths and their performance. While the shortest path (red) has the highest efficiency score Enav = 0.86, excessive collisions with the movable objects degrade the effort score Inav = 0.56. In contrast, while the longest path (blue) is collision-free Inav = 1.00, efficiency is largely compromised Enav = 0.54. A balanced path (yellow) potentially offers the best trade-off. C. Metrics … view at source ↗
Figure 6
Figure 6. Figure 6: Three teleoperated paths for the area clearing task, where [PITH_FULL_IMAGE:figures/full_fig_p006_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

28 extracted references · 6 linked inside Pith

  1. [1]

    Navigation among movable obstacles: Real-time reasoning in complex environments,

    M. Stilman and J. J. Kuffner, “Navigation among movable obstacles: Real-time reasoning in complex environments,”International Journal of Humanoid Robotics, vol. 2, no. 04, pp. 479–503, 2005

    2005

  2. [2]

    Probabilistic mobile manipulation in dynamic environments, with application to opening doors

    A. Petrovskaya and A. Y . Ng, “Probabilistic mobile manipulation in dynamic environments, with application to opening doors.” inIJCAI, 2007, pp. 2178–2184

    2007

  3. [3]

    Planning and executing navigation among movable obstacles,

    M. Stilman, K. Nishiwaki, S. Kagami, and J. J. Kuffner, “Planning and executing navigation among movable obstacles,”Advanced Robotics, vol. 21, no. 14, pp. 1617–1634, 2007

    2007

  4. [4]

    Local path planning among pushable objects based on reinforcement learning,

    L. Yao, V . Modugno, A. M. Delfaki, Y . Liu, D. Stoyanov, and D. Kanoulas, “Local path planning among pushable objects based on reinforcement learning,” inIEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 3062–3068

    2024

  5. [5]

    Spatial action maps for mobile manipulation,

    J. Wu, X. Sun, A. Zeng, S. Song, J. Lee, S. Rusinkiewicz, and T. Funkhouser, “Spatial action maps for mobile manipulation,”arXiv preprint arXiv:2004.09141, 2020

    Pith/arXiv arXiv 2004

  6. [6]

    Pushing through clutter with movability awareness of blocking obstacles,

    J. J. Weeda, S. Bakker, G. Chen, and J. Alonso-Mora, “Pushing through clutter with movability awareness of blocking obstacles,” arXiv preprint arXiv:2502.20106, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    The ycb object and model set: Towards common benchmarks for manipulation research,

    B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar, “The ycb object and model set: Towards common benchmarks for manipulation research,” inIEEE International Conference on Advanced Robotics, 2015, pp. 510–517

    2015

  8. [8]

    robosuite: A modular simulation framework and benchmark for robot learning,

    Y . Zhu, J. Wong, A. Mandlekar, R. Mart´ın-Mart´ın, A. Joshi, S. Nasiri- any, and Y . Zhu, “robosuite: A modular simulation framework and benchmark for robot learning,”arXiv preprint arXiv:2009.12293, 2020

    Pith/arXiv arXiv 2009

  9. [9]

    Interactive gibson benchmark: A benchmark for interactive navigation in cluttered environments,

    F. Xia, W. B. Shen, C. Li, P. Kasimbeg, M. E. Tchapmi, A. Toshev, R. Mart´ın-Mart´ın, and S. Savarese, “Interactive gibson benchmark: A benchmark for interactive navigation in cluttered environments,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 713–720, 2020

    2020

  10. [10]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” inIEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 5026–5033

    2012

  11. [11]

    Pymunk: A easy-to-use pythonic rigid body 2d physics library. (version 6.7.0),

    V . Blomqvist, “Pymunk: A easy-to-use pythonic rigid body 2d physics library. (version 6.7.0),” 2024. [Online]. Available: https://pymunk.org

    2024

  12. [12]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inInternational conference on machine learning. PMLR, 2018, pp. 1861–1870

    2018

  13. [13]

    Proximal policy optimization algorithms,

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

    Pith/arXiv arXiv 2017

  14. [14]

    Auto-icenav: A local navigation strategy for autonomous surface ships in broken ice fields,

    R. de Schaetzen, A. Botros, N. Zhong, K. Murrant, R. Gash, and S. L. Smith, “Auto-icenav: A local navigation strategy for autonomous surface ships in broken ice fields,”arXiv preprint arXiv:2411.17155, 2024

    arXiv 2024

  15. [15]

    Autonomous navigation in ice-covered waters with learned predictions on ship-ice interactions,

    N. Zhong, A. Potenza, and S. L. Smith, “Autonomous navigation in ice-covered waters with learned predictions on ship-ice interactions,” inIEEE International Conference on Robotics and Automation, 2025, pp. 10 157–10 163

    2025

  16. [16]

    Interactive-far: Interactive, fast and adaptable routing for navigation among movable obstacles in complex unknown environments,

    B. He, G. Chen, W. Wang, J. Zhang, C. Fermuller, and Y . Aloimonos, “Interactive-far: Interactive, fast and adaptable routing for navigation among movable obstacles in complex unknown environments,” in IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024, pp. 5402–5409

    2024

  17. [17]

    Search-based path planning in interactive environments among movable obstacles,

    Z. Ren, B. Suvonov, G. Chen, B. He, Y . Liao, C. Fermuller, and J. Zhang, “Search-based path planning in interactive environments among movable obstacles,”arXiv preprint arXiv:2410.18333, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Unwieldy object delivery with nonholonomic mobile base: A stable pushing approach,

    Y . Tang, H. Zhu, S. Potters, M. Wisse, and W. Pan, “Unwieldy object delivery with nonholonomic mobile base: A stable pushing approach,” IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7727–7734, 2023

    2023

  19. [19]

    Unwieldy object delivery with nonholonomic mobile base: A free pushing approach,

    Y . Tang, M. Wisse, and W. Pan, “Unwieldy object delivery with nonholonomic mobile base: A free pushing approach,”IEEE Robotics and Automation Letters, 2024

    2024

  20. [20]

    Autonomous multiple-trolley collection system with nonholonomic robots: Design, control, and implementation,

    P. Xie, B. Xia, A. Hu, Z. Zhao, L. Meng, Z. Sun, X. Gao, J. Wang, and M. Q.-H. Meng, “Autonomous multiple-trolley collection system with nonholonomic robots: Design, control, and implementation,”Journal of Field Robotics, vol. 42, no. 1, pp. 20–36, 2025

    2025

  21. [21]

    Benchmarking metric ground navigation,

    D. Perille, A. Truong, X. Xiao, and P. Stone, “Benchmarking metric ground navigation,” in2020 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR). IEEE, 2020, pp. 116–121

    2020

  22. [22]

    Benchmarking reinforcement learning techniques for autonomous navigation,

    Z. Xu, B. Liu, X. Xiao, A. Nair, and P. Stone, “Benchmarking reinforcement learning techniques for autonomous navigation,” in IEEE International Conference on Robotics and Automation, 2023, pp. 9224–9230

    2023

  23. [23]

    Arena-bench: A benchmarking suite for obstacle avoidance approaches in highly dynamic environments,

    L. K ¨astner, T. Bhuiyan, T. A. Le, E. Treis, J. Cox, B. Meinardus, J. Kmiecik, R. Carstens, D. Pichel, B. Fatlounet al., “Arena-bench: A benchmarking suite for obstacle avoidance approaches in highly dynamic environments,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 9477–9484, 2022

    2022

  24. [24]

    Gymnasium: A standard interface for reinforcement learning environments,

    M. Towers, A. Kwiatkowski, J. Terry, J. U. Balis, G. De Cola, T. Deleu, M. Goul ˜ao, A. Kallinteris, M. Krimmel, A. KGet al., “Gymnasium: A standard interface for reinforcement learning environments,”arXiv preprint arXiv:2407.17032, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Randomized Kinodynamic Planning,

    S. M. LaValle and J. J. Kuffner, “Randomized Kinodynamic Planning,” The International Journal of Robotics Research, vol. 20, no. 5, pp. 378–400, May 2001

    2001

  26. [26]

    Streamlining object pushing: Behavior tree-based coordination of control and planning,

    F. Bertoncelli and L. Sabattini, “Streamlining object pushing: Behavior tree-based coordination of control and planning,” inIEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems, 2024, pp. 5203– 5210

    2024

  27. [27]

    Stable-baselines3: Reliable reinforcement learning im- plementations,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning im- plementations,”Journal of Machine Learning Research, vol. 22, no. 268, pp. 1–8, 2021

    2021

  28. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inIEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016