Pith. sign in

REVIEW 3 major objections 6 minor 30 references

An end-to-end RL-to-spiking pipeline deploys warehouse AGV pathfinding on a neuromorphic chip at up to 11,281× lower energy and roughly half the latency of a GPU ANN while matching decision quality.

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 · grok-4.5

2026-07-12 13:17 UTC pith:TQOUGB2U

load-bearing objection Solid full-stack offline demo of DQN→SNN→Speck for warehouse pathfinding with real energy numbers; the closed-loop gap is real but does not erase the result. the 3 major comments →

arxiv 2606.20031 v3 pith:TQOUGB2U submitted 2026-06-18 cs.RO cs.AI

A Neuromorphic Reinforcement Learning Framework for Efficient Pathfinding in Robotic Mobile Fulfillment Systems

Junzhe Xu , Zecui Zeng , Lusong Li , Yuetong Fang , Renjing Xu This is my paper
classification cs.RO cs.AI
keywords neuromorphic computingspiking neural networksreinforcement learningRMFSANN-to-SNN conversionAGV pathfindingenergy-efficient robotics
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 establishes that a trained deep Q-network policy for multi-AGV pathfinding in robotic mobile fulfillment systems can be converted into a spiking network and run on physical neuromorphic hardware without collapsing decision quality. Conventional search methods scale poorly under dynamic multi-agent constraints, and ordinary deep RL is too power-hungry for edge robots. The authors first train the ANN with a collision-allowing strategy that densifies useful trajectories in narrow warehouse aisles, then apply hard-label distillation and first-layer weight scaling so the SNN still selects the correct action at only four time-steps. Hardware measurements on the neuromorphic chip report up to 11,281 times lower energy per decision and nearly half the latency of the same policy on a high-end GPU, with conversion rates near 1.0 and high task-completion rates for one to eight agents. A reader who cares about practical warehouse robotics therefore sees a concrete route to large fleets that remain real-time and battery-viable.

Core claim

SDQN-RMFS shows that an RL pathfinding policy trained as a full-precision ANN can be converted, via collision-allowing pretraining plus hard-label knowledge distillation and parameter scaling, into an SNN that runs on a neuromorphic chip at four time-steps while preserving essentially the same action choices, delivering measured energy reductions of up to 11,281 imes and roughly twofold lower latency relative to the original ANN on an RTX 4090 for multi-AGV warehouse tasks.

What carries the argument

Hard-label knowledge distillation: the trained ANN’s continuous Q-values are turned into one-hot pseudo-labels by argmax and used to fine-tune a student network so its output margins become robust to the residual membrane-potential quantization that appears at ultra-low time-steps; first-layer weight scaling further guarantees early spikes so information reaches the output within the latency budget.

Load-bearing premise

The energy, latency and quality numbers rest on offline sequential replay of pre-recorded state trajectories into the chip, plus an external safety rule that can force a stop, being a faithful proxy for true closed-loop multi-AGV control.

What would settle it

Place the same converted SNN on the neuromorphic chip in closed-loop control of 1–8 live AGVs inside the warehouse simulator or physical testbed, log real energy, end-to-end latency, completion degree and how often the safety override fires under simultaneous agent motion and sensor noise; if energy savings fall below roughly 1,000 imes or completion drops well below the original ANN, the central claim fails.

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

If this is right

  • Large multi-AGV fleets can execute continuous pathfinding under tight edge power budgets without continuous GPU-class compute.
  • Inference latency for warehouse decisions can be cut roughly in half relative to a high-end GPU while energy falls by four to five orders of magnitude.
  • Conversion fidelity at four time-steps stays high enough that external collision overrides remain rare fallbacks rather than the primary controller.
  • The same collision-allowing pretraining densifies informative trajectories for other constrained multi-agent navigation domains.

Where Pith is reading between the lines

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

  • Offline trajectory replay may understate closed-loop effects such as simultaneous multi-agent state changes or sensor noise; online chip-in-the-loop experiments are the direct next measurement.
  • Hard-label distillation that widens action margins rather than preserving exact Q-values could transfer to other discrete value-based robotic control problems where ranking noise is the main conversion risk.
  • If neuromorphic inference becomes standard on AGVs, warehouse energy accounting would shift from compute-dominated to motion-dominated, changing battery and thermal design assumptions.

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

3 major / 6 minor

Summary. The paper proposes SDQN-RMFS, an end-to-end pipeline for multi-AGV pathfinding in Robotic Mobile Fulfillment Systems: a Double DQN is trained with a collision-allowing exploration strategy and sparse binary state maps, then converted to an IF-based SNN via first-layer weight scaling and hard-label (argmax) knowledge distillation of the Q-head, and finally deployed offline on the SPECK2E neuromorphic chip. In a 16 imes16 warehouse simulation with 1–8 AGVs, the converted SNN at 4 time-steps reaches conversion rates near 1.0 and high completion degrees (Table II). Hardware profiling by replaying recorded trajectories reports up to 11,281× lower energy and roughly half the latency versus the original ANN on an RTX 4090 (Table III), while an external collision-detection safety layer can override actions to ‘stop’.

Significance. If the reported energy/latency numbers and near-lossless policy transfer hold under realistic operation, the work would be a concrete step toward edge-deployable RL for warehouse robotics, where prior RL–SNN studies largely remain in simulation. Strengths include: (i) a hardware-aware conversion recipe (scaling + hard-label distillation + bias removal) that demonstrably enables T=4 inference with CR≈1.0 (Fig. 6, Table II); (ii) an ablation of the collision limit CL that improves multi-AGV training stability (Fig. 4); and (iii) actual physical measurements on SPECK2E rather than simulated spike counts alone. These are useful engineering contributions for neuromorphic robotics even if claims must be scoped more carefully to the offline protocol used.

major comments (3)
  1. §V-E and Table III measure energy (0.89–1.95×10^{-3} mJ) and latency (0.061–0.079 ms) by offline sequential replay of pre-recorded state trajectories on Speck; the chip never closes the loop with live multi-AGV dynamics, timing, or sensor noise. The abstract and §VI nevertheless claim a “practical and energy-sustainable pathway for large-scale RMFS operations.” Offline open-loop inference cost does not automatically transfer to online closed-loop energy, effective latency, or decision quality. Either add closed-loop chip-in-the-loop experiments (or a clear online proxy with membrane-state continuity and inter-agent timing) or substantially narrow the abstract/conclusion claims to offline inference cost under the stated protocol.
  2. §V-A.3 introduces an external action-detection safety layer that forces ‘stop’ when a collision is judged likely; the text states that high CR “minimizes” overrides but never reports override frequency or an ablation without the safety filter. Table II completion degrees and action lengths are therefore not pure measures of the converted policy. Please report the fraction of steps overridden for each (AGVs, k, distillation) setting and, if feasible, completion degree with the safety layer disabled, so readers can separate policy fidelity from the engineering fallback.
  3. §IV-B–C and §V-A.1 formulate multi-AGV pathfinding with other agents’ positions in the state, yet training and evaluation appear to use independent (or shared) DQN policies without explicit communication or joint action selection. Table II shows strong completion for 8 AGVs, but it is unclear whether deadlocks, livelocks, or priority conflicts arise and how often the safety layer resolves them. Clarify the multi-agent training protocol (independent learners vs. parameter sharing, simultaneous vs. sequential decisions) and add basic coordination metrics (e.g., deadlock rate, mean waiting time) so the multi-AGV claims are interpretable.
minor comments (6)
  1. §IV-D.1: the claim that removing biases (b_l=0) “preserves functional equivalence” should be supported by a short ANN-with-bias vs. ANN-without-bias comparison before conversion, or at least by stating the performance drop on the teacher.
  2. Fig. 4 and Fig. 5: axis labels and units (especially action length vs. episode) are hard to read in the manuscript text; ensure high-resolution figures with explicit legends for CL and algorithm variants.
  3. Table I: the last layer is listed as 5(4)×1×1; briefly restate in the caption when the action space is 4 vs. 5 to avoid ambiguity.
  4. §II-B: several RL–SNN navigation works are cited; a short explicit contrast (simulation-only vs. physical chip, continuous vs. discrete actions, time-steps used) would better position the SPECK2E deployment contribution.
  5. Eq. (2): the shaped reward uses pos_n and pos_l; define pos_n in the surrounding text as clearly as pos_c/pos_l/pos_t in §IV-B.
  6. Typographical: “Von Neumann” appears as “V on Neumann” in §V-E.1; “SPECK2E” vs. “Speck” should be consistent.

Circularity Check

0 steps flagged

No circular derivation chain; energy/latency claims and conversion fidelity are empirical measurements and ablations, not identities forced by fitted inputs or self-citation.

full rationale

The paper's load-bearing results (Table II conversion rates/completion degrees under distillation+scaling; Table III Speck vs. RTX-4090 energy/latency) are obtained by direct measurement after training and conversion, not by algebraic reduction to the free parameters (CL, k, T, reward weights) that are merely ablated. Hard-label distillation (argmax of teacher Q-values as one-hot targets for the student) is an explicit engineering step to sharpen the output distribution before IF conversion; it does not redefine the target metric by construction, nor does parameter scaling (W1_snn = k·W1_ann) force the reported CR=1.0 or the 11 281 imes energy ratio. Self-citations appear only as background (e.g., prior SNN conversion or Speck hardware descriptions) and are not invoked as uniqueness theorems that forbid alternatives or that alone underwrite the central hardware claims. Offline trajectory replay is a methodological limitation of the evaluation protocol, not a circular step in any derivation. The pipeline is therefore self-contained against its own empirical benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The claim rests on standard MDP/DQN and IF-neuron conversion math, plus several hand-chosen training and conversion knobs (collision limit, first-layer scale k, time-steps T=4, shaped rewards) and the modeling choice that offline trajectory replay plus a rule-based safety filter adequately represents edge RMFS inference. No new physical entities are postulated; the ‘invented’ items are algorithmic constructs of the pipeline.

free parameters (4)
  • collision limit CL = 40
    Hand-chosen training hyperparameter; ablated in Fig. 4 and fixed at CL=40 for main results. Directly affects trajectory density and final policy.
  • first-layer weight scale k = 7 (hardware)
    Hardware-aware free scale applied as W1_snn = k·W1_ann to force early spikes; selected via Table II (k=7 used for hardware).
  • SNN time-steps T = 4
    Inference latency/energy knob; main claims use T=4 after distillation+scaling.
  • shaped reward magnitudes (Rs, Rgoal, Rfail) = e.g. ±0.1, −0.2, −0.5, Rgoal/Rfail unspecified numeric
    Hand-designed dense rewards in Eq. (2) and surrounding text; not learned; drive exploration and completion metrics.
axioms (4)
  • domain assumption IF-neuron rate approximates ReLU with residual error vl(t)/t that vanishes as T grows (Eq. 3).
    Standard ANN-to-SNN conversion premise used throughout §IV-D; paper then works at ultra-low T where residual is large.
  • domain assumption RMFS pathfinding is adequately modeled as a discrete MDP with binary map stack state and 4–5 actions.
    §IV-B problem formulation; continuous dynamics, communication, and pod physics are abstracted away.
  • ad hoc to paper Argmax hard labels from the teacher Q-network define the correct policy for distillation (cross-entropy student).
    §IV-D.2; trades relative Q-value structure for one-hot margin; assumed sufficient for navigation.
  • ad hoc to paper Removing biases (bl=0) preserves functional equivalence for this architecture on Speck.
    §IV-D.1 hardware alignment choice; not generally true for arbitrary ANNs.
invented entities (2)
  • SDQN-RMFS full-stack pipeline no independent evidence
    purpose: Name the combined collision-allowing DQN + hard-label distillation + Speck deployment system.
    Organizational construct; not a physical entity. Independent evidence is the reported experiments themselves.
  • Hard-label knowledge distillation for RL Q-heads before SNN conversion no independent evidence
    purpose: Force one-hot-like outputs so low-T IF quantization does not flip argmax actions.
    Methodological invention relative to classification-oriented conversion literature; validated only inside this paper’s tables.

pith-pipeline@v1.1.0-grok45 · 17097 in / 3517 out tokens · 40553 ms · 2026-07-12T13:17:40.381146+00:00 · methodology

0 comments
read the original abstract

Dynamic environmental changes, confined workspaces, and stringent real-time constraints make pathfinding in Robotic Mobile Fulfillment Systems (RMFS) a challenging problem for conventional search- and rule-based methods, which typically suffer from high computational complexity and long decision latency. While reinforcement learning (RL) has emerged as a powerful alternative, deploying learned policies with extreme energy efficiency on resource-constrained hardware remains an open challenge. We present SDQN-RMFS, an end-to-end framework that achieves high-fidelity deployment of an RL-trained policy from a full-precision artificial neural network (ANN) through to a neuromorphic chip. By computing only when triggered by sparse events, this framework unlocks ultra-low-power RMFS pathfinding. Our full-stack pipeline operates as follows: an ANN policy is first efficiently trained via a collision-allowing strategy to densify informative trajectories, and then converted into a spiking neural network (SNN) via a hard-label knowledge distillation approach. This effectively addresses the output distribution mismatch, preserving policy capability across the ANN-to-SNN pipeline while substantially reducing inference latency. Hardware experiments demonstrate up to 11,281$\times$ energy savings and a nearly two-fold reduction in latency compared to a high-performance GPU baseline, while maintaining decision quality on par with the original trained policy. These results establish physical neuromorphic inference as a practical and energy-sustainable pathway for large-scale RMFS operations.

Figures

Figures reproduced from arXiv: 2606.20031 by Junzhe Xu, Lusong Li, Renjing Xu, Yuetong Fang, Zecui Zeng.

Figure 1
Figure 1. Figure 1: (a) Sketch map of RMFS, which includes several AGVs with or [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The three-stage process of the SDQN-RMFS framework. First, an ANN agent is trained using RL to interact with the RMFS, collecting state-action [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The process of distillation before converting it to an SNN to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The impact of different CL on the DQN’s training process for various numbers of AGVs. The figures show the completion degree, which is the ratio of successfully transported pods. The dashed lines represent the smoothing of the completion degree and action length curves [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of different standard RL baselines such as Dueling [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

Review history (2 revisions) →

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

30 extracted references · 2 linked inside Pith

  1. [1]

    Robotic mobile ful- fillment systems: A survey on recent developments and research opportunities,

    ´I. R. da Costa Barros and T. P. Nascimento, “Robotic mobile ful- fillment systems: A survey on recent developments and research opportunities,”Robotics and Autonomous Systems, vol. 137, p. 103729, 2021

    2021

  2. [2]

    Primal: Pathfinding via reinforcement and imitation multi-agent learning,

    G. Sartoretti, J. Kerr, Y . Shi, G. Wagner, T. S. Kumar, S. Koenig, and H. Choset, “Primal: Pathfinding via reinforcement and imitation multi-agent learning,”IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2378–2385, 2019

    2019

  3. [3]

    Primal 2: Pathfind- ing via reinforcement and imitation multi-agent learning-lifelong,

    M. Damani, Z. Luo, E. Wenzel, and G. Sartoretti, “Primal 2: Pathfind- ing via reinforcement and imitation multi-agent learning-lifelong,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2666–2673, 2021

    2021

  4. [4]

    E. W. Dijkstra,A Note on Two Problems in Connexion with Graphs, 1st ed. New York, NY , USA: Association for Computing Machinery, 2022, p. 287–290. [Online]. Available: https://doi.org/10. 1145/3544585.3544600

    arXiv 2022

  5. [5]

    A formal basis for the heuristic determination of minimum cost paths,

    P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,”IEEE transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968

    1968

  6. [6]

    Path planning for multiple mobile robots in smart warehouse,

    A. Bolu and ¨O. Korc ¸ak, “Path planning for multiple mobile robots in smart warehouse,” in2019 7th International Conference on Control, Mechatronics and Automation (ICCMA). IEEE, 2019, pp. 144–150

    2019

  7. [7]

    Integrated task assignment and path planning for capacitated multi- agent pickup and delivery,

    Z. Chen, J. Alonso-Mora, X. Bai, D. D. Harabor, and P. J. Stuckey, “Integrated task assignment and path planning for capacitated multi- agent pickup and delivery,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 5816–5823, 2021

    2021

  8. [8]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al., “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  9. [9]

    Acquisition of automated guided vehicle route planning policy using deep reinforcement learning,

    R. Kamoshida and Y . Kazama, “Acquisition of automated guided vehicle route planning policy using deep reinforcement learning,” in 2017 6th IEEE international conference on advanced logistics and transport (ICALT). IEEE, 2017, pp. 1–6

    2017

  10. [10]

    Deep reinforcement learning: A brief survey,

    K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, “Deep reinforcement learning: A brief survey,”IEEE signal processing magazine, vol. 34, no. 6, pp. 26–38, 2017

    2017

  11. [11]

    A* guiding dqn algorithm for automated guided vehicle pathfinding problem of robotic mobile fulfillment systems,

    L. Luo, N. Zhao, Y . Zhu, and Y . Sun, “A* guiding dqn algorithm for automated guided vehicle pathfinding problem of robotic mobile fulfillment systems,”Computers & Industrial Engineering, vol. 178, p. 109112, 2023

    2023

  12. [12]

    Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless navi- gation with neuromorphic hardware,

    G. Tang, N. Kumar, and K. P. Michmizos, “Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless navi- gation with neuromorphic hardware,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 6090– 6097

    2020

  13. [13]

    Toward energy-efficient spike-based deep reinforcement learning with temporal coding,

    M. Zhang, S. Wang, J. Wu, W. Wei, D. Zhang, Z. Zhou, S. Wang, F. Zhang, and Y . Yang, “Toward energy-efficient spike-based deep reinforcement learning with temporal coding,”IEEE Computational Intelligence Magazine, vol. 20, no. 2, pp. 45–57, 2025

    2025

  14. [14]

    Networks of spiking neurons: the third generation of neural network models,

    W. Maass, “Networks of spiking neurons: the third generation of neural network models,”Neural networks, vol. 10, no. 9, pp. 1659–1671, 1997

    1997

  15. [15]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain,et al., “Loihi: A neuromorphic manycore processor with on-chip learning,”Ieee Micro, vol. 38, no. 1, pp. 82–99, 2018

    2018

  16. [16]

    Improved robustness of reinforcement learning policies upon conver- sion to spiking neuronal network platforms applied to atari breakout game,

    D. Patel, H. Hazan, D. J. Saunders, H. T. Siegelmann, and R. Kozma, “Improved robustness of reinforcement learning policies upon conver- sion to spiking neuronal network platforms applied to atari breakout game,”Neural Networks, vol. 120, pp. 108–115, 2019

    2019

  17. [17]

    A neuro-inspired approach to intelligent collision avoidance and navigation,

    N. Salvatore, S. Mian, C. Abidi, and A. D. George, “A neuro-inspired approach to intelligent collision avoidance and navigation,” in2020 AIAA/IEEE 39th Digital Avionics Systems Conference (DASC). IEEE, 2020, pp. 1–9

    2020

  18. [18]

    Speck: A smart event-based vision sensor with a low latency 327k neuron convolutional neuronal network processing pipeline,

    O. Richter, Y . Xing, M. De Marchi, C. Nielsen, M. Katsimpris, R. Cattaneo, Y . Ren, Y . Hu, Q. Liu, S. Sheik,et al., “Speck: A smart event-based vision sensor with a low latency 327k neuron convolutional neuronal network processing pipeline,”arXiv preprint arXiv:2304.06793, 2023

    arXiv 2023

  19. [19]

    Strategy and benchmark for con- verting deep q-networks to event-driven spiking neural networks,

    W. Tan, D. Patel, and R. Kozma, “Strategy and benchmark for con- verting deep q-networks to event-driven spiking neural networks,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 11, 2021, pp. 9816–9824

    2021

  20. [20]

    Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing,

    P. U. Diehl, D. Neil, J. Binas, M. Cook, S.-C. Liu, and M. Pfeiffer, “Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing,” in2015 International joint conference on neural networks (IJCNN). ieee, 2015, pp. 1–8

    2015

  21. [21]

    Optimal conversion of conventional artificial neural networks to spiking neural networks,

    S. Deng and S. Gu, “Optimal conversion of conventional artificial neural networks to spiking neural networks,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=FZ1oTwcXchK

    2021

  22. [22]

    Optimal ANN-SNN conversion for high-accuracy and ultra-low- latency spiking neural networks,

    T. Bu, W. Fang, J. Ding, P. DAI, Z. Yu, and T. Huang, “Optimal ANN-SNN conversion for high-accuracy and ultra-low- latency spiking neural networks,” inInternational Conference on Learning Representations, 2022. [Online]. Available: https: //openreview.net/forum?id=7B3IJMM1k M

    2022

  23. [23]

    Error-aware conversion from ann to snn via post-training parameter calibration,

    Y . Li, S. Deng, X. Dong, and S. Gu, “Error-aware conversion from ann to snn via post-training parameter calibration,”International Journal of Computer Vision, vol. 132, no. 9, pp. 3586–3609, 2024

    2024

  24. [24]

    Masked spiking transformer,

    Z. Wang, Y . Fang, J. Cao, Q. Zhang, Z. Wang, and R. Xu, “Masked spiking transformer,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 1761–1771

    2023

  25. [25]

    Deep reinforcement learning with double q-learning,

    H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double q-learning,” inProceedings of the AAAI conference on artificial intelligence, vol. 30, no. 1, 2016

    2016

  26. [26]

    Differential coding for training-free ann-to-snn conversion,

    Z. Huang, W. Fang, T. Bu, P. Xue, Z. Hao, W. Liu, Y . Tang, Z. Yu, and T. Huang, “Differential coding for training-free ann-to-snn conversion,”arXiv preprint arXiv:2503.00301, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    A free lunch from ann: Towards efficient, accurate spiking neural networks calibration,

    Y . Li, S. Deng, X. Dong, R. Gong, and S. Gu, “A free lunch from ann: Towards efficient, accurate spiking neural networks calibration,” inInternational conference on machine learning. PMLR, 2021, pp. 6316–6325

    2021

  28. [28]

    Dueling network architectures for deep reinforcement learning,

    Z. Wang, T. Schaul, M. Hessel, H. Hasselt, M. Lanctot, and N. Freitas, “Dueling network architectures for deep reinforcement learning,” in International conference on machine learning. PMLR, 2016, pp. 1995–2003

    2016

  29. [29]

    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

  30. [30]

    Actor-critic algorithms,

    V . Konda and J. Tsitsiklis, “Actor-critic algorithms,”Advances in neural information processing systems, vol. 12, 1999

    1999