{"id":"25aa90ed-4b63-4977-b3d6-f74be9f6b7a6","arxiv_id":"2607.17317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An RL-based scheduler chooses CPU/GPU/DLA clusters for each perception model to meet scene-derived frame-rate targets, reporting large energy savings, though the frame-rate targets are defined by a hand-tuned entropy mapping.","lead":"TAPAS is a runtime scheduling framework that adjusts the frame rate of an autonomous vehicle's perception module based on scene complexity, then assigns each neural network to the best available compute cluster to meet that frame rate at low energy. It reports 93–100% throughput met rate with up to 76% energy savings on KITTI and 97%/64% on unseen nuScenes, but the target frame rates come from a hand-tuned entropy mapping rather than from independent safety requirements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Entropy-derived FPS targets are self-referential: throughput met rate validates TAPAS against targets generated by hand-tuned Eq. (4), not against any independent safety or downstream planning requirement.","rationale":"The reader identified the same load-bearing weakness: the entropy-to-FPS mapping is hand-tuned and never validated against an independent safety or planning metric. I agree. The paper's strongest claim depends on the correctness of these self-generated throughput targets; without external grounding, the reported 93–100% throughput met rate and energy savings may simply reflect under-provisioning relative to arbitrary targets. The paper is otherwise coherent, with a substantial evaluation and no obvious internal mathematical error. The issue is addressable: a closed-loop planner-based evaluation would settle whether the entropy-derived targets are safe. Given the reader already issued a CONDITIONAL verdict, my recommendation is UNCHANGED — the concern reinforces the condition rather than overturning the verdict.","tokens_in":19699,"tokens_out":3571,"duration_ms":39886,"concrete_test":"Re-run the evaluation in a closed-loop simulator (e.g., CARLA or Gazebo/ROS 2) with a planner/controller consuming TAPAS's variable-FPS perception outputs. Compare collision rate, distance-to-collision, and planning error against fixed 5/10/15 FPS policies over KITTI and nuScenes scenarios. Concretely, group frames by entropy bin; for low-entropy (5 FPS) frames under high ego-vehicle speed, test whether planner performance degrades. Use the same perception models and Jetson Orin NX hardware, running at least ~20 episodes per condition. If the entropy-derived targets yield collision rates statistically indistinguishable from fixed 15 FPS, the concern is resolved; if not, the reported energy savings are not safety-preserving.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — 93–100% throughput met rate with 76% energy savings — is measured against FPS targets produced by the paper's own entropy-to-FPS mapping (Eqs. 2–4, Sec. III-C1), with hand-tuned parameters H_base=1.5, CG_h=1.0, T_base=10, CG_t=5, N_h=N_t=3 (Sec. IV-A). The mapping assumes scene complexity (number/classes of detected objects) directly determines the correct perception rate. No independent evidence is provided that these targets preserve downstream planning or safety. Indeed, Sec. II-A1 explicitly decouples perception from safety-critical reasoning, and the evaluation never feeds TAPAS's variable FPS into a planner or a safety metric. A sparse highway at high speed can be dangerous at 5 FPS despite low entropy; a dense but stopped queue can be safe at 10 FPS despite high entropy. Thus 'throughput met rate' is largely self-referential: it rewards hitting a target whose validity is the very thing at issue. This is not an internal inconsistency — the framework may do exactly what it claims — but it means the headline result may not correspond to operational safety requirements. The missing TANGO comparison and missing code are secondary; the unvalidated target mapping is the load-bearing concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":20077,"tokens_out":4466,"duration_ms":50493,"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":[{"comment":"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.","section":"Section III-C1, Eqs. (3)–(4) and Section IV-A"},{"comment":"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.","section":"Section III-B, Eq. (1)"},{"comment":"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.","section":"Section III-C1, Eq. (2)"}],"minor_comments":[{"comment":"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.","section":"Algorithm 2, Line 5"},{"comment":"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.","section":"Section IV-B, Figure 11"},{"comment":"The paper reports PPO training but omits training hyperparameters (learning rate, discount factor, number of episodes, convergence criteria, etc.). Please provide these for reproducibility.","section":"Section IV-A"},{"comment":"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.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The systems contribution is real, but the main evaluation metric is circular because the FPS targets come from the authors' own hand-tuned mapping. The sum-of-FPS issue in Eq. (1) is a technical error that may change the optimization and results. I would like the authors to validate the entropy-to-FPS mapping against an external criterion or substantially weaken the claims, and to fix the aggregate-throughput definition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: TAPAS is a real systems paper with a genuine new combination — scene-entropy-derived FPS targets plus RL-based dynamic model-to-cluster mapping for multi-DNN perception pipelines on HMPs. If you accept the upstream assumption about what the \"right\" FPS target is, the measured 76% energy savings and 93–100% met rate on KITTI, and 64%/97% on nuScenes, are meaningful and evaluated on real hardware. The problem is exactly where the stress test points: the targets are generated by the paper's own hand-tuned Eq. (4) with H_base=1.5, T_base=10, CG_t=5, and \"throughput met rate\" is measured against those same targets. That makes the headline metric partially self-referential. The paper never checks against a downstream planner or a safety metric, and a low-entropy highway at speed could still demand more than 5 FPS. So the load-bearing weakness is the target-generation step, not the scheduler.\n\nWhat is actually good: the coupling is new compared to Roborun/SHIFT (which adapt model settings) and TANGO/OmniBoost/Band (which schedule at fixed FPS). The ablation of GRU+RRM is useful, and the overhead analysis (2.3 ms, ~45 mJ, <0.2% remapping overhead) is credible. The cross-dataset generalization to nuScenes adds confidence that the entropy-to-FPS mapping is at least stable in distribution.\n\nThe other soft spots: Equation (1) sums per-model FPS to represent aggregate throughput, which is questionable if the target is a module-level FPS — a pipeline running four models at 10 FPS each has an aggregate of 40 but still processes frames at 10. That formulation needs a fix or a clarification. The missing TANGO baseline is odd because it's from the same group and is the closest multi-DNN scheduler on HMPs. There are no error bars anywhere, and no code or traces are released. The RRM details are thin; it's a Qwen2-based reward model with \"grounding,\" but the training data and the grounding loss are not described.\n\nOverall: it's a coherent, honest paper, no sign of fabrication, and the central scheduling contribution likely works as described. The target-generation issue is addressable — either validate against an independent safety/planning metric, or explicitly frame the contribution as meeting the system's own specified FPS targets. I would send it to peer review and expect heavy revision, not rejection. It's the kind of paper a reading group would enjoy picking apart, and I'd cite it for the scheduling/energy results.","headline":"Solid systems paper whose headline numbers are weakened by self-generated FPS targets; the scheduling piece is real and worth referee time.","tokens_in":20538,"tokens_out":3293,"would_cite":true,"duration_ms":35899,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Scene-driven frame rates meet perception targets 93-100% of the time","keywords":["throughput-adaptive perception","spatial entropy","reinforcement learning","heterogeneous multi-core processors","energy efficiency","autonomous systems","model-to-cluster mapping","GRU"],"falsifier":"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.","tokens_in":19618,"feed_emoji":"🤖","tokens_out":2977,"duration_ms":28681,"temperature":0.7,"pith_summary":"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.","feed_headline":"Scene-aware frame rates cut robot perception energy by 76%","feed_subtitle":"A self-tuning scheduler meets its own perception targets 93-100% of the time on KITTI and generalizes to unseen data.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Adaptive perception trims robot energy by 76%","Scene-aware FPS scheduling cuts energy 76%","TAPAS: Self-tuning perception meets targets, saves energy","Perception that adapts to scene complexity cuts power","Dynamic FPS control reduces perception energy by 76%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive perception trims robot energy by 76%","Scene-aware FPS scheduling cuts energy 76%","TAPAS: Self-tuning perception meets targets, saves energy","Perception that adapts to scene complexity cuts power","Dynamic FPS control reduces perception energy by 76%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1166,"prompt_tokens":779,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":523,"tokens_out":387,"duration_ms":4590,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:20:03.399065+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}