Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Scene-driven frame rates meet perception targets 93-100% of the time

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 18:20 UTC pith:TSUZK5OE

load-bearing objection Solid systems paper whose headline numbers are weakened by self-generated FPS targets; the scheduling piece is real and worth referee time. the 3 major comments →

arxiv 2607.17317 v1 pith:TSUZK5OE submitted 2026-07-19 cs.LG cs.AI

TAPAS: Throughput-adaptive Perception for Autonomous Systems

Aman Vyas , Vasista Kodumagulla , Zain Taufique , Pasi Liljeberg , Anil Kanduri This is my paper
classification cs.LG cs.AI
keywords throughput-adaptive perceptionspatial entropyreinforcement learningheterogeneous multi-core processorsenergy efficiencyautonomous systemsmodel-to-cluster mappingGRU
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.

Self-driving cars, drones, and robots rely on perception pipelines whose compute needs change with scene complexity. Existing systems run at a fixed frame rate, wasting energy in simple scenes or falling behind in complex ones. This paper claims that a module computing spatial entropy from detection outputs can estimate the right frame rate for each scene, and that a reinforcement-learned scheduler can then assign each model in the perception pipeline to the best available CPU/GPU/DLA cluster to meet that target at minimum energy. On a Jetson Orin NX, the resulting system reportedly meets its own throughput targets 93-100% of the time on KITTI test sequences and 97% on unseen nuScenes, while cutting energy by up to 76% and 64% respectively. If these numbers hold, adaptive throughput could become a practical way to stretch battery life on mobile autonomous systems.

Core claim

The central claim is that perception throughput should be adapted at runtime to scene complexity rather than fixed at design time. TAPAS estimates scene complexity via Shannon entropy computed over the semantic class map from an object detector, maps entropy bands to FPS targets (5/10/15), and trains a GRU-based reinforcement learning agent to assign each perception model to a compute cluster so the achieved FPS meets the target while energy is minimized. The agent is trained offline on KITTI profiles using PPO with a grounded reward reasoning model, then deploys with minimal overhead at runtime. The paper reports 93-100% throughput met rate with energy savings up to 76% on KITTI, and 97% me

What carries the argument

Spatial entropy: a Shannon entropy computed from the object detector's class map, used as a scene-complexity proxy that sets the frame-rate target via a hand-tuned piecewise mapping. The GRU-based RL agent: a gated recurrent unit that takes temporal state (entropy, FPS target, workload features) and outputs cluster assignments for each perception model; trained with PPO. The reward reasoning model (RRM): a grounded reward generator that combines measured throughput/energy with entropy-driven context instead of a fixed heuristic.

Load-bearing premise

The paper assumes that a frame rate derived from spatial entropy, via hand-set parameters, is the correct throughput requirement for safe operation, but this mapping is never validated against any independent safety or planning metric.

What would settle it

A human or automated audit could run a downstream collision-avoidance planner with (a) TAPAS's entropy-derived frame rates and (b) a fixed 10 FPS baseline, on KITTI/nuScenes; if the fixed baseline achieves equal or better safety with similar or lower energy, the claim that entropy-driven targets improve efficiency without safety loss would be falsified.

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

If this is right

  • Fixed-FPS perception strategies over-provision simple scenes and under-provision complex ones; scene-aware FPS targets eliminate both, so energy scales with actual demand.
  • Because the scheduler only remaps when the FPS target changes (threshold Δ), runtime overhead stays below 0.2% and the policy adds only ms-level latency.
  • The entropy-to-FPS estimator transfers to unseen datasets (nuScenes) without retraining, suggesting scene complexity is a portable signal across driving environments.
  • The agent degrades gracefully under hardware unavailability: GPU failures in the highest-demand region drop throughput from 100% to 86% at 25% unavailability, and CPU/DLA failures are mostly benign.

Where Pith is reading between the lines

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

  • If entropy-based FPS targets are taken as given, the method's reported met rates are somewhat circular; the real test is whether downstream planning performance (e.g., collision avoidance) is preserved at the lowered frame rates.
  • The hand-tuned mapping parameters (H_base, CG_h, T_base, CG_t, N_h, N_t) could be learned or validated against safety outcomes; the paper leaves this open.
  • The framework is presented as model-agnostic middleware; it could be extended to other heterogeneous SoCs and to include DVFS or model approximation as additional actions for finer energy granularity.

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 / 4 minor

Summary. TAPAS proposes a runtime framework for adaptive perception on heterogeneous mobile/edge platforms. It uses spatial entropy from object-detection outputs to estimate variable FPS targets, and a GRU-based PPO agent with a Reward Reasoning Model to map perception models to CPU/GPU/DLA clusters so that the estimated target FPS is met with low energy. The authors evaluate on Jetson Orin NX using KITTI and nuScenes, reporting 93–100% throughput met rate with up to 76% energy savings on KITTI, and 97% throughput met rate with 64% lower energy on nuScenes compared to EE, OmniBoost, and Band baselines.

Significance. If the central claims hold, TAPAS addresses a real problem: fixed-FPS perception pipelines waste energy on heterogeneous edge platforms. The paper's strengths include deployment on actual hardware, evaluation on an unseen dataset, a recurrent RL agent with grounded reward modeling, and ablations comparing agent architectures and reward schemes. The result would be a practical advance for energy-efficient autonomous perception. However, the headline metric 'throughput met rate' is measured against FPS targets that the authors themselves define from entropy thresholds; without external validation against downstream planning or safety requirements, the reported gains are not interpretable as operational improvements. The formulation of aggregate throughput in Eq. (1) also appears technically incorrect, which further undermines the optimization and evaluation.

major comments (3)
  1. [Section III-C1, Eqs. (3)–(4) and Section IV-A] The headline metric 'throughput met rate' is measured against FPS targets generated by the paper's own entropy-to-FPS mapping. The parameters (H_base=1.5, CG_h=1.0, T_base=10, CG_t=5, N_h=N_t=3) are hand-tuned on KITTI. There is no independent validation that these targets preserve safety or downstream planning performance. Section II-A1 explicitly decouples perception from safety-critical reasoning, yet the evaluation never feeds TAPAS's variable FPS into a planner or a safety metric. A low-entropy highway scene at high speed may require more FPS than a dense but stopped queue; the mapping is not self-evident. Thus the reported 93–100% met rate and energy savings do not establish operational utility. I request an evaluation against a downstream planning/safety metric (e.g., Zhuyi's safety-based rate estimation) or a clear reframing of the claim as meeting self-defined targets.
  2. [Section III-B, Eq. (1)] The aggregate achieved throughput is defined as the sum over per-model rates R_i(h_{i,t}). For a perception module with N concurrent tasks, the system's frame-processing rate is not the sum of per-model FPS. If four models each process frames at 5 FPS, the pipeline is not delivering 20 FPS; it is delivering four parallel streams at 5 FPS. A correct constraint should be min_i R_i >= T_v (or an equivalent pipeline-level frame rate). As written, the objective can be minimized by increasing per-model rates on easy tasks while starving a critical task, and the reported throughput met rate may be inflated. This issue affects the RL reward, the deployment policy, and the comparisons in Figures 16 and 20.
  3. [Section III-C1, Eq. (2)] The quantity called 'spatial entropy' is the Shannon entropy of the histogram of class labels in the detection output, H_t = -Σ p_t(c) log p_t(c). This is not a spatial entropy; it ignores the spatial arrangement of objects. The monotonic relationship with the number of detected objects (Figure 14b) is an expected property of histogram entropy, not evidence that this proxy captures scene complexity relevant to required perception rate. If this proxy is the sole basis for FPS targets, it requires validation against an independent notion of scene complexity or task-relevant risk.
minor comments (4)
  1. [Algorithm 2, Line 5] The call Ψ(I_t, O, I_t, T_base, ...) appears to have an extra I_t argument compared to Algorithm 1. Please correct the inconsistency.
  2. [Section IV-B, Figure 11] The agent ablation names 'ANN [18]', but Reference [18] is the TANGO scheduler. Clarify whether the ANN baseline is TANGO or a generic MLP. TANGO is not included in the SOTA comparisons in Figures 16/20; including it would strengthen the evaluation.
  3. [Section IV-A] The paper reports PPO training but omits training hyperparameters (learning rate, discount factor, number of episodes, convergence criteria, etc.). Please provide these for reproducibility.
  4. [Section IV-B] The text says 'we emulate variable throughput by skipping frames' and also claims runtime adaptation on hardware. Clarify whether the evaluation uses actual frame skipping or whether the scheduler changes execution speed; this affects the validity of the energy measurements.

Circularity Check

3 steps flagged

Entropy-to-FPS mapping is self-definitional; throughput met rate validates TAPAS against targets generated by the paper's own hand-tuned Eq. (4).

specific steps
  1. self definitional [Section III-C1, Eqs. (3)-(4); Section IV-C(a)/(c), Fig. 13(a)]
    "Figure 13 (a) and (b) present the robustness of the throughput estimator on unseen nuScenes dataset. Figure (a) shows the relationship between scene complexity and the average FPS selected by TAPAS across four nuScenes sequences. As spatial entropy increases from 0.5 (Seq 1) to 2.0 (Seq 4), the average FPS correspondingly rises from approximately 8 FPS to 12 FPS, confirming its ability to match processing rates to the complexity of unseen scenes."

    Eq. (4) defines each FPS target T_vi = T_base + (j - ceil(Nt/2))*CG_t as a deterministic function of the entropy band [H_k, H_{k+1}) defined in Eq. (3), with hand-set parameters. Higher entropy therefore necessarily maps to higher or equal FPS by construction. The observed rise from ~8 to ~12 FPS is a consequence of the definition, not an empirical confirmation. Calling this 'confirming its ability to match processing rates' is circular; there is no independent measure of what FPS is actually required.

  2. self definitional [Algorithm 1/Line 4 and Section III-C5; Section IV-A; Abstract]
    "throughput estimator Ψ(·) is invoked to compute variable FPS target T v with arguments (T base, Nh, Nt, CGt, CGh) (Line 4). These serve as ground-truth supervisory signals for variable throughput scheduling. ... For evaluation, we set throughput estimator parameters as N h = N t = 3, H base = 1.5, CG h = 1.0, T base = 10FPS, and CG t = 5FPS."

    The targets called 'ground-truth supervisory signals' are produced by the paper's own throughput estimator, not by any independent safety, planning, or downstream task metric. Since throughput met rate is measured against these self-generated targets, the headline 93-100% / 97% met rate means TAPAS meets targets that TAPAS itself defined through hand-tuned Eq. (4). The correctness of the entropy-to-FPS mapping is assumed rather than validated.

  3. self definitional [Section IV-C(d), Fig. 15(b)]
    "Figure 15(b) presents joint entropy-FPS density contours reveal a strong correlation between scene complexity and assigned FPS targets. ... The consistent entropy–FPS ordering across both datasets validates the estimator’s ability to generalize the learned mapping."

    The entropy-FPS 'correlation' and 'ordering' are identities: Eq. (4) assigns FPS deterministically from entropy bands, and Eq. (3) sets those bands from entropy. A deterministic monotone mapping cannot fail to produce a strong entropy-FPS correlation. Treating this as validation of a 'learned mapping' is circular; the only non-circular observation is that nuScenes has a different entropy distribution, which does not validate the FPS targets themselves.

full rationale

The central circularity is in the throughput-target half of the paper. TAPAS defines FPS targets by its own Eqs. (3)-(4) with hand-tuned parameters, then measures throughput met rate against those targets and presents the entropy-FPS correlation as validation. This makes the headline throughput-met-rate claim self-referential: the scheduler is rewarded for hitting targets the paper itself generated, and no independent safety or planning metric checks whether those targets are correct. The entropy-to-FPS ordering shown in Figs. 13 and 15 is guaranteed by the definitional mapping, so it is not independent evidence. However, the GRU+PPO scheduling policy, the real-hardware energy measurements, and the comparisons to fixed-FPS baselines are meaningful engineering contributions that do not reduce to the target-definition issue; if one accepts the paper's target standard, the scheduler's energy savings have independent content. No load-bearing self-citation chain was found. The score of 6 reflects partial circularity in the central metric and estimator validation, not a fully circular derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The central functional dependencies rest on the hand-tuned throughput estimator and on assumptions about the relationship between entropy, required FPS, and aggregate throughput that are not independently grounded.

free parameters (6)
  • H_base = 1.5
    Baseline spatial entropy for threshold mapping, hand-set; directly sets FPS target levels via Eq. 3-4.
  • CG_h = 1.0
    Entropy class-gap spacing; controls sensitivity of FPS target to entropy.
  • T_base = 10 FPS
    Baseline FPS target; chosen as average perception rate; centers FPS levels.
  • CG_t = 5 FPS
    FPS step size; together with N_t fixes targets {5,10,15}.
  • N_h = N_t = 3 = 3
    Number of entropy/FPS levels; chosen to simplify demonstration; not derived from requirements.
  • Delta (remap threshold) = unspecified
    Algorithm 2 line 7 requires a threshold Δ to trigger remapping; not quantified anywhere, though it affects scheduling overhead and met rate.
axioms (4)
  • domain assumption More complex scenes (higher entropy) require higher perception FPS.
    Introduced in Sec III-C1 and motivated by [5], [6]; not independently validated for these pipelines; the paper only shows its own mapping produces this result.
  • domain assumption Spatial entropy of object detection class histogram is a sufficient proxy for scene complexity.
    Equation 2 computes a histogram entropy over detection class labels; the paper does not compare to other complexity metrics or downstream task performance.
  • ad hoc to paper Sum of per-model achieved FPS (Eq. 1) represents aggregate perception throughput.
    Equation 1 defines the objective with a sum over workloads; physically, concurrent tasks on one camera stream have a pipeline FPS bounded by the bottleneck, not the sum. If sum is used, throughput constraint may be trivially satisfied.
  • domain assumption Frame skipping faithfully emulates low-FPS perception.
    Sec IV-B.a adopts frame skipping from [31] to emulate variable FPS; this assumes discarded frames do not alter required perception behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 19425 in / 12490 out tokens · 118377 ms · 2026-08-01T18:20:03.399065+00:00 · methodology

0 comments
read the original abstract

Autonomous systems rely on a perception module to navigate through dynamic environments. In real-world scenarios, the perception module's throughput requirements vary at runtime due to changes in scene complexity. However, existing perception strategies assume a fixed FPS and static model-to-cluster mapping, resulting in either over/under provision of throughput requirements or unnecessary energy consumption across diverse scenes. Addressing this challenge requires tightly coupled \textit{scene complexity awareness} to estimate an appropriate FPS target and \textit{dynamic model-to-cluster mapping} to deliver the required throughput at minimum energy. We propose a throughput-adaptive perception strategy for mobile/edge platforms, enabling intelligent runtime resource allocation based on varying FPS targets. We use Reinforcement Learning (RL) with RRM (Reward Reasoning Model) and a GRU (Gated Recurrent Unit) agent to orchestrate perception tasks across heterogeneous mobile/edge platforms. We evaluate TAPAS on Jetson Orin NX across KITTI and unseen nuScenes. On the \textit{KITTI} dataset's test sequences, TAPAS achieves 93-100% throughput met rate while saving energy by 76%. On the unseen \textit{nuScenes} dataset, TAPAS maintains 97% throughput met rate with 64% lower energy compared to \textit{SOTA} approaches, proving its robustness.

Figures

Figures reproduced from arXiv: 2607.17317 by Aman Vyas, Anil Kanduri, Pasi Liljeberg, Vasista Kodumagulla, Zain Taufique.

Figure 1
Figure 1. Figure 1: Overview of static and adaptive perception modules: SOTA [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Impact of hardware vs. algorithmic knobs on energy and accuracy [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Distribution of FPS targets across KITTI sequences. (b) Mean [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a). Region-wise FPS targets in KITTI Sequence 7, (b). Normalized [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Energy comparison under fixed (5, 10, and 15 FPS) and variable [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training the GRU-based RL agent with offline profiled data generated using [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: RL-based reward shaping approaches: heuristic-based reward, [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison between KITTI and nuScenes datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Sequence-wise spatial entropy and average detected objects [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: GT vs. Predicted entropy KDE and cross-dataset joint density. [PITH_FULL_IMAGE:figures/full_fig_p011_15.png] view at source ↗
Figure 17
Figure 17. Figure 17: Impact of cluster availability on normalized energy gain [PITH_FULL_IMAGE:figures/full_fig_p012_17.png] view at source ↗
Figure 20
Figure 20. Figure 20: (a) and (b) jointly illustrate normalized energy consumption and throughput met rate for fixed 5/10/15 FPS targets with EE, OmniBoost, and Band, alongside the adap￾tive TAPAS strategy, across different mix combinations. EE exhibits linear energy scaling (2× at 10 FPS, 3× at 15 FPS) due to fixed FPS processing and static mapping, resulting in 69–90% higher energy than TAPAS and poor throughput met rate (< … view at source ↗
Figure 19
Figure 19. Figure 19: Normalized energy per sequence over different strategies. [PITH_FULL_IMAGE:figures/full_fig_p013_19.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

36 extracted references · 8 linked inside Pith

  1. [1]

    Semantically-driven deep reinforcement learning for inspection path planning,

    G. Malczyk, M. Kulkarni, and K. Alexis, “Semantically-driven deep reinforcement learning for inspection path planning,”IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 7206–7213, 2025

    2025

  2. [2]

    Transformer-based model for monocular visual odometry: A video understanding approach,

    A. O. Franc ¸ani and M. R. O. A. Maximo, “Transformer-based model for monocular visual odometry: A video understanding approach,”IEEE Access, vol. 13, p. 13959–13971, 2025

    2025

  3. [3]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthineret al., “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https: //arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2021

  4. [4]

    Roboshape: Using topology patterns to scalably and flexibly deploy accelerators across robots,

    S. M. Neuman, R. Ghosal, T. Bourgeat, B. Plancher, and V . J. Reddi, “Roboshape: Using topology patterns to scalably and flexibly deploy accelerators across robots,” inProc. of Int. Symp. on Computer Archi- tecture, ser. ISCA ’23, 2023

    2023

  5. [5]

    Zhuyi: perception processing rate estimation for safety in autonomous vehicles,

    Y .-S. Hsiao, S. K. S. Hari, M. Filipiuk, T. Tsai, M. B. Sullivan, V . J. Reddi, V . Singh, and S. W. Keckler, “Zhuyi: perception processing rate estimation for safety in autonomous vehicles,” inProc. of ACM/IEEE Design Automation Conf. (DAC), 2022, pp. 289–294

    2022

  6. [6]

    How fast is too fast? the role of perception latency in high-speed sense and avoid,

    D. Falanga, S. Kim, and D. Scaramuzza, “How fast is too fast? the role of perception latency in high-speed sense and avoid,”IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 1884–1891, 2019

    2019

  7. [7]

    Driving scenario perception-aware computing system design in autonomous vehicles,

    H. Zhao, Y . Zhang, P. Meng, H. Shi, L. E. Li, T. Louet al., “Driving scenario perception-aware computing system design in autonomous vehicles,” inProc. of IEEE Int. Conf. on Computer Design (ICCD), 2020, pp. 88–95

    2020

  8. [8]

    Roborun: A robot runtime to exploit spatial heterogeneity,

    B. Boroujerdian, R. Ghosal, J. Cruz, B. Plancher, and V . J. Reddi, “Roborun: A robot runtime to exploit spatial heterogeneity,” inProc. of ACM/IEEE Design Automation Conf. (DAC), 2022

    2022

  9. [9]

    Context-aware multi-model object detection for diversely heterogeneous compute systems,

    J. Davis and M. E. Belviranli, “Context-aware multi-model object detection for diversely heterogeneous compute systems,” inProc. of IEEE Design, Automation & Test in Europe Conf. & Exhibition (DATE), 2024, pp. 1–6

    2024

  10. [10]

    Working with dla,

    N. Corporation, “Working with dla,” 2025. [On- line]. Available: https://docs.nvidia.com/deeplearning/tensorrt/latest/ inference-library/work-with-dla.html

    2025

  11. [11]

    Energy-efficient mobile robot control via run-time monitoring of envi- ronmental complexity and computing workload,

    S. A. Mohamed, M.-H. Haghbayan, A. Miele, O. Mutlu, and J. Plosila, “Energy-efficient mobile robot control via run-time monitoring of envi- ronmental complexity and computing workload,” inProc. of IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2021, pp. 7587– 7593

    2021

  12. [12]

    A coordinated approach to control mechanical and computing resources in mobile robots,

    S. Shahsavari, H. Haghbayan, A. Miele, E. Immonen, and J. Plosila, “A coordinated approach to control mechanical and computing resources in mobile robots,”IEEE Trans. on Robotics, vol. 41, pp. 347–363, 2025

    2025

  13. [13]

    Band: coordinated multi-dnn inference on heterogeneous mobile processors,

    J. S. Jeong, J. Lee, D. Kim, C. Jeon, C. Jeong, Y . Leeet al., “Band: coordinated multi-dnn inference on heterogeneous mobile processors,” inProc. of Int. Conf. on Mobile Systems, Applications and Services (MobiSys), 2022, p. 235–247

    2022

  14. [14]

    Omniboost: Boosting throughput of heterogeneous embedded devices under multi-dnn workload,

    A. Karatzas and I. Anagnostopoulos, “Omniboost: Boosting throughput of heterogeneous embedded devices under multi-dnn workload,” inProc. of ACM/IEEE Design Automation Conf. (DAC), 2023, pp. 1–6

    2023

  15. [15]

    Axonn: Energy-aware execution of neural network inference on multi-accelerator heterogeneous socs,

    I. Dagli, A. Cieslewicz, J. McClurg, and M. E. Belviranli, “Axonn: Energy-aware execution of neural network inference on multi-accelerator heterogeneous socs,” inProc. of ACM/IEEE Design Automation Con- ference, 2022, pp. 1069–1074

    2022

  16. [16]

    Equipping diffusion models with differentiable spatial entropy for low-light image enhancement,

    W. Lian, W. Lian, and Z. Luo, “Equipping diffusion models with differentiable spatial entropy for low-light image enhancement,” in Proc. of IEEE/CVF Conf. on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024, pp. 6671–6681

    2024

  17. [17]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017. [Online]. Available: https://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  18. [18]

    Tango: Low latency multi-dnn inference on heterogeneous edge platforms,

    Z. Taufique, A. Vyas, A. Miele, P. Liljeberg, and A. Kanduri, “Tango: Low latency multi-dnn inference on heterogeneous edge platforms,” in Proc. of Int. Conf. on Computer Design (ICCD), 2024, pp. 300–307

    2024

  19. [19]

    Reward reasoning model,

    J. Guo, Z. Chi, L. Dong, Q. Dong, X. Wu, S. Huang, and F. Wei, “Reward reasoning model,” 2025. [Online]. Available: https://arxiv.org/abs/2505.14674

    Pith/arXiv arXiv 2025

  20. [20]

    Continuous, real-time object detection on mobile devices without offloading,

    M. Liu, X. Ding, and W. Du, “Continuous, real-time object detection on mobile devices without offloading,” inProc. of IEEE Int. Conf. on Distributed Computing Systems (ICDCS), 2020, pp. 976–986

    2020

  21. [21]

    Maritime vessel tank inspection using aerial robots: Experience from the field and dataset release,

    M. Dharmadhikari, N. Khedekar, P. D. Petris, M. Kulkarni, M. Nissov, and K. Alexis, “Maritime vessel tank inspection using aerial robots: Experience from the field and dataset release,” 2024. [Online]. Available: https://arxiv.org/abs/2404.19045

    Pith/arXiv arXiv 2024

  22. [22]

    The use of spatial information in entropy measures,

    L. Altieri, D. Cocchi, and G. Roli, “The use of spatial information in entropy measures,” 2017. [Online]. Available: https://arxiv.org/abs/ 1703.06001

    Pith/arXiv arXiv 2017

  23. [23]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” 2014. [Online]. Available: https://arxiv.org/abs/1412.3555

    Pith/arXiv arXiv 2014

  24. [24]

    Unpack- ing reward shaping: understanding the benefits of reward engineering on sample complexity,

    A. Gupta, A. Pacchiano, Y . Zhai, S. M. Kakade, and S. Levine, “Unpack- ing reward shaping: understanding the benefits of reward engineering on sample complexity,” inProc. of Int. Conf. on Neural Information Processing Systems (NeurIPS), 2022

    2022

  25. [25]

    Reward design with language models,

    M. Kwon, S. M. Xie, K. Bullard, and D. Sadigh, “Reward design with language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2303.00001

    Pith/arXiv arXiv 2023

  26. [26]

    Qwen2 technical report,

    A. Yanget al., “Qwen2 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2407.10671

    Pith/arXiv arXiv 2024

  27. [27]

    Survey of hallucination in natural language generation,

    Z. Jiet al., “Survey of hallucination in natural language generation,” ACM Computing Surveys, vol. 55, no. 12, p. 1–38, Mar. 2023

    2023

  28. [28]

    Nvidia jetson Orin,

    NVIDIA, “Nvidia jetson Orin,” https://www.nvidia.com/en-us/ autonomous-machines/embedded-systems/jetson-orin/, 2024

    2024

  29. [29]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,”The International Journal of Robotics Research, vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  30. [30]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Kr- ishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” 2020

    2020

  31. [31]

    Slimslam: An adaptive runtime for visual-inertial simultaneous localization and mapping,

    A. Behroozi, Y . Chen, V . Fruchter, L. Subramanian, S. Srikanth, and S. Mahlke, “Slimslam: An adaptive runtime for visual-inertial simultaneous localization and mapping,” inProc. of ACM Int. Conf. 15 on Architectural Support for Programming Languages and Operating Systems (ASPLOS), Volume 3, 2024, p. 900–915

    2024

  32. [32]

    A deep reinforcement learning based online scheduling policy for deep neural network multi-tenant multi-accelerator systems,

    F. G. Blancoet al., “A deep reinforcement learning based online scheduling policy for deep neural network multi-tenant multi-accelerator systems,” inProc. of ACM/IEEE Design Automation Conf. (DAC), 2024

    2024

  33. [33]

    Deepvo: Towards end-to- end visual odometry with deep recurrent convolutional neural networks,

    S. Wang, R. Clark, H. Wen, and N. Trigoni, “Deepvo: Towards end-to- end visual odometry with deep recurrent convolutional neural networks,” inProc. of IEEE Int. Conf. on Robotics and Automation (ICRA), 2017, p. 2043–2050

    2017

  34. [34]

    Suraksha: A framework to analyze the safety implications of perception design choices in avs,

    H. Zhao and et al, “Suraksha: A framework to analyze the safety implications of perception design choices in avs,” inProc. of IEEE Int. Symp. on Software Reliability Engineering (ISSRE), 2021, pp. 434–445

    2021

  35. [35]

    Frugal following: power thrifty object detection and tracking for mobile augmented reality,

    K. Apicharttrisorn, X. Ran, J. Chen, S. V . Krishnamurthy, and A. K. Roy-Chowdhury, “Frugal following: power thrifty object detection and tracking for mobile augmented reality,” inProc. of ACM Conf. on Embedded Networked Sensor Systems (SenSys), 2019, p. 96–109

    2019

  36. [36]

    Adaptive optimization of autonomous vehicle computational resources for performance and energy improve- ment,

    S. Jambotkar, L. Guo, and Y . Jia, “Adaptive optimization of autonomous vehicle computational resources for performance and energy improve- ment,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 7594–7600

    2021