Thermal-Headroom Routing for Efficient Robot Perception in Extreme Heat

Sparsh Roy; Kevin Zhou; Arhaan Shah; Shreyan Garg; Vihaan Joshi-Apte; Maxim Kudryashov; Nanda Cheruvanthala

cs.ROOct 26, 2026

Thermal-Headroom Routing for Efficient Robot Perception in Extreme Heat

Authors: Sparsh Roy, Kevin Zhou, Arhaan Shah, Shreyan Garg, Vihaan Joshi-Apte, Maxim Kudryashov, Nanda Cheruvanthala

Organizations: Massachusetts Institute of Technology, Cambridge, MA, USA · Hopewell Valley Central High School, Pennington, NJ, USA · Paradise Valley High School, Phoenix, AZ, USA

Abstract

Adaptive robot perception systems must balance detector accuracy, inference latency, and energy consumption, particularly when high ambient temperatures cause processors to thermally throttle. We propose Thermal-Headroom Routing (THR), a lightweight model-selection policy that forecasts processor temperature using a closed-form RC thermal model and excludes detector configurations predicted to cross a throttling threshold. We evaluate THR with four YOLO11 configurations across seven real-world perception traces, eight ambient temperatures, and 20 randomized seeds, comparing it with latency-aware routing, fixed-threshold thermostats, duty caps, and a matched-compute-budget control. At ambient temperatures of 45-55 degrees C and a tight inference deadline, THR eliminates throttling in simulation, reduces deadline misses by 3.9-7.5 percentage points, and reduces energy per on-time frame by approximately 12-13% relative to a latency-belief router. A separate experiment using deadline-paced, nonoverlapping frame execution yields miss-rate reductions of 6.2-9.5 percentage points. A tuned thermostat or duty cap can match THR at a fixed operating point; predictive routing is most useful when ambient conditions or calibrated thermal parameters differ from those used to tune static policies. Measurements on an Apple M5 support the single-node heating approximation but do not validate hardware throttling. The results characterize the conditions under which thermal forecasting improves the efficiency and timeliness of resource-constrained robot perception.

Figures & tables

Fig. 1: SoC temperature during one 2-minute mission at 45 °C (fog trace, seed 0). Static C1 and Belief go through repeated throttle cycles. THR alternates C1 and C2 and stays just below the release threshold.

$\Delta t$ (frame)	0.1 s	$P_{\mathrm{idle}}$ / $P_{\max}$	5 / 25 W
$\tau$	30 s	$T_{\mathrm{hi}}$ / $T_{\mathrm{lo}}$	90 / 85 °C
$R_{\mathrm{th}}$	2.0 °C/W	throttled power	$0.6\times$

TABLE I: Thermal-plant parameters (fixed before any run).

Fig. 2: Ambient sweep (E2), base deadline. Lines are the mean over seven traces; bands span the lowest and highest trace. Left: deadline-miss rate. Middle: share of the mission throttled. Right: longest run of consecutive misses (log scale). THR (red) and the thermostat never throttle. The cap tuned at 45 °C wastes accuracy below 45 °C (lower miss, Table IV ) and throttles above it.

Ambient 45 °C
Policy	Miss	Thr.	Blackout	$P$	$E_{\mathrm{ot}}$	Acc.
Static C1	60.3	20.3	12.5	23.0	5.79	1.000
Static C4	0.0	0.0	0.0	5.6	0.56	0.695
Reactive	44.2	12.4	0.9	22.6	4.06	0.975
Belief	44.3	12.2	0.9	22.6	4.07	0.975
Thermostat	41.2	0.0	0.9	22.2	3.79	0.958

TABLE II: Efficiency, quality and cost in E2 (base deadline), mean over 7 traces and 20 seeds. Miss and Thr. (throttled) in %. Blackout is the longest run of consecutive misses (s).

P

is mean SoC power (W) and

E_{\mathrm{ot}}

is energy per on-time frame (J). Acc. is pseudo-GT agreement.

	Ambient 45 °C						Ambient 55 °C
Trace	Blackout C1	Thr. Belief	Thr. THR	$\Delta$ THR	THR $-$ Tst	THR $-$ MB	Blackout C1	Thr. Belief	Thr. THR	$\Delta$ THR	THR $-$ Tst	THR $-$ MB
Off-road (clean)	12.5	13	0	+4.9 ∗	+1.6 ∗	+0.2	90.7	38	0	+7.2 ∗	+3.9 ∗	-0.1
Off-road seg.	12.5	12	0	+4.8 ∗	+1.6 ∗	-0.2	90.7	37	0	+6.8 ∗	+3.8 ∗	-0.3
Rain	12.5	13	0	+4.4 ∗	+1.9 ∗	-0.1	90.8	36	0	+7.1 ∗	+3.7 ∗	-0.3
Snow	12.5	12	0	+5.4 ∗	+1.7 ∗	+0.1	90.7	38	0	+7.2 ∗	+4.1 ∗	-0.1
Fog	12.4	11	0	+4.8 ∗	+1.9 ∗	+0.0	90.7	34	0	+7.5 ∗	+3.8 ∗	+0.0

TABLE III: Heat regime (E2) per trace, base deadline. Blackout C1 is Static C1’s longest run of consecutive misses (s). Thr. is the % of the mission throttled.

\Delta

THR is the reduction in miss rate relative to Belief (pp). THR

-

Tst and THR

-

MB are THR’s miss-rate advantage over the 88 °C thermostat and the matched-budget control (positive means THR misses less). ∗ : Holm-significant across the seven traces.

Fig. 3: Accuracy against deadline-miss rate (E2 and E9), mean over seven traces. The grey line is a fixed C1 cap between 0.2 and 1.0. Every temperature-aware policy lies on it; Belief sits off the curve at 55 °C because throttling raises its miss rate.

	Cap tuned at 45 °C			THR
Ambient	Thr.	Miss	Acc.	Thr.	Miss	Acc.
35 °C	0.0	39.5	0.946	0.0	49.9	1.000
40 °C	0.0	39.5	0.946	0.0	46.5	0.982
45 °C	0.0	39.5	0.946	0.0	39.6	0.950
50 °C	15.6	36.9	0.930	0.0	32.3	0.914
55 °C	31.0	31.2	0.902	0.0	25.1	0.883

TABLE IV: A C1 cap tuned once at 45 °C against THR at other ambients (mean over traces). Thr. and Miss in %.

	Ambient 45 °C					Ambient 55 °C
True chip	$\Delta$ THR	Sig.	Belief	Tst.	THR	$\Delta$ THR	Sig.	Belief	Tst.	THR
$R_{\mathrm{th}}\times 0.8$	+0.0 to +0.0	0/7	0.0	0.0	0.0	+4.7 to +6.1	7/7	17.0	0.0	0.0
$R_{\mathrm{th}}\times 1.2$	+6.7 to +7.9	7/7	28.4	0.0	0.0	+8.2 to +12.1	7/7	51.8	29.0	0.0
$\tau\times 0.5$	+7.9 to +9.6	7/7	15.8	0.0	0.0	+11.0 to +14.7	7/7	42.3	2.7	0.0
$\tau\times 2$	+3.7 to +4.4	7/7	0.0	0.0	0.0	+4.8 to +5.6	7/7	25.5	0.0	0.0
throttle power $0.8\times$	+3.4 to +5.3	7/7	12.8	0.0	0.0	+4.8 to +6.7	7/7	41.0	0.0	0.0

TABLE V: A different true chip (E6). THR uses the matching model.

\Delta

THR: reduction in miss rate relative to Belief (pp, range over traces) and Holm-significant traces. Max thr.: largest throttled share over traces (%) for Belief, the 88 °C thermostat and THR.

	Thr. (%)					THR vs
Amb.	B	Tst	Cap	THR	$\Delta$ (pp)	B	Tst	MB	Cap
35 °C	0.0	0.0	0.0	0.0	+0.0 to +0.1	0/0	0/7	0/0	0/7
45 °C	15.5	0.0	0.0	0.0	+6.2 to +7.8	7/0	7/0	0/0	0/0
50 °C	27.6	0.0	23.0	0.0	+8.2 to +9.5	7/0	7/0	0/0	7/0
55 °C	40.3	1.5	37.4	0.0	+8.0 to +9.1	7/0	7/0	0/0	7/0

TABLE VI: E11, consistent frame timing (one decision per

D

), mean over traces. Thr.: % of the mission throttled for Belief (B), the 88 °C thermostat (Tst), the cap tuned at 45 °C and THR.

\Delta

: THR’s miss reduction relative to Belief (pp, range over traces). Last four columns: traces where THR is Holm-significantly better / worse than each policy on miss rate (MB: matched budget).

		Miss (%)			Throttled (%)			U			THR vs B
$D$	Amb.	B	Tst	THR	B	Tst	THR	B	Tst	THR	miss	U
1.0 $\times$	35 °C	49.9	49.9	49.9	0.0	0.0	0.0	0.501	0.501	0.501	0/0	0/0
	45 °C	44.3	41.2	39.6	12.2	0.0	0.0	0.532	0.546	0.554	7/0	5/0
	55 °C	32.3	29.0	25.1	36.3	0.0	0.0	0.607	0.610	0.632	7/0	6/0
1.25 $\times$	35 °C	0.0	0.0	0.0	0.0	0.0	0.0	1.000	1.000	1.000	0/0	0/0
	45 °C	0.5	0.0	0.0	12.2	0.0	0.0	0.970	0.958	0.950	7/0	1/6

TABLE VII: Deadline sweep (E5): means over 7 traces for Belief (B), the 88 °C thermostat (Tst) and THR. U is on-time utility. The last two columns count traces where THR is Holm-significantly better / worse than Belief on miss rate and on U.

Fig. 4: Heat plus fault-coupled contention at 45 °C (E3): reduction in miss rate relative to Belief, 95% CIs over 20 seeds.

	25 °C		45 °C		55 °C
Trace	J	J+T	J	J+T	J	J+T
Off-road (clean)	+0.0	+0.0	+0.0	+6.0 ∗	+0.0	+7.9 ∗
Off-road seg.	+12.7 ∗	+12.7 ∗	+12.2 ∗	+13.1 ∗	+9.0 ∗	+13.3 ∗
Rain	+9.7 ∗	+9.7 ∗	+9.5 ∗	+9.7 ∗	+5.8 ∗	+10.1 ∗
Snow	+5.3 ∗	+5.3 ∗	+3.5 ∗	+7.0 ∗	+1.6 ∗	+8.8 ∗
Fog	+12.5 ∗	+12.5 ∗	+8.8 ∗	+9.1 ∗	+3.4 ∗	+9.2 ∗

TABLE VIII: E3: reduction in miss rate relative to Belief (pp). J is Joint; J+T is Joint + THR. ∗ : Holm-significant.

THR’s model	$\Delta$ miss (pp)	Sig.	Acc.
true model	+3.9 to +5.4	7/7	0.917 to 0.994
$R_{\mathrm{th}}\times 0.7$	+0.8 to +1.8	7/7	0.943 to 0.996
$R_{\mathrm{th}}\times 1.3$	+7.1 to +9.0	7/7	0.888 to 0.993
$\tau\times 0.5$	+5.6 to +7.6	7/7	0.899 to 0.994
$\tau\times 2$	+3.2 to +4.4	7/7	0.925 to 0.995

TABLE IX: THR with a wrong thermal model at 45 °C (E4): range of the miss reduction relative to Belief, Holm-significant traces, accuracy range and the largest throttled share.

	Burst share (%)			Max run
Trace	Belief	Joint	$\Delta$ (pp)	Belief	Joint
Off-road (clean)	18.9	18.9	+0.0	7.6	7.6
Off-road seg.	17.4	14.7	+2.8 ∗	8.4	7.3
Rain	14.3	13.7	+0.6 ∗	6.7	6.7
Snow	20.0	17.4	+2.6 ∗	7.6	7.3
Fog	21.9	16.3	+5.6 ∗	7.1	6.8

TABLE X: Bursts under fault-coupled contention (E1). Burst share: % of frames in runs of at least 3 misses. Max run in frames.

Cfg	Lat.	First	Last	$\Delta P$	$\tau$	$\Delta T$	$R$	RMSE
C1	189	187	186	14.6	20	41.1	2.81	0.99
C2	55	55	55	13.8	19	38.9	2.82	0.80
C4	7	7	7	7.7	45	25.7	3.33	0.88

TABLE XI: Apple M5 under nonstop YOLO11 inference. Lat.: median latency over the run, its first 30 s and its last 30 s (ms).

\Delta P

: GPU+CPU power above idle (W).

\tau

and

\Delta T

: fitted time constant (s) and steady-state rise above idle ( °C).

R=\Delta T/\Delta P

(°C/W). RMSE of the fit ( °C).

Fig. 5: Apple M5 GPU temperature under nonstop YOLO11 inference (thin) with the fitted one-node RC step response (thick).

Power-Efficient Inference

See all 16 papers

Jul 31, 2026

Belief-Space Perception Routing under Coupled Sensor Faults and Compute Contention

A robot that has to see and react on a fixed clock runs into two problems at once. Its cameras degrade in rain, mud, fog, and darkness. And the single onboard processor it runs on is shared with planning and control, so the compute left over for perception moves around from second to second. Most systems model the two separately. We present a perception router that tracks probabilistic estimates of sensor-fault state and compute- contention state, couples them with a noisy-OR term, and uses the coupled estimate to pick one of four detector configurations (YOLO11x/n at 1280 or 640 px) so that the frame finishes before its deadline. Where the two stressors co-occur, the coupled policy cuts the deadline-miss rate by 1.1 to 9.4 percentage points against a policy that treats them independently. The interval excludes zero in five of six conditions, the pooled effect over 10 sequences and 6 conditions has sign-test p = 0.001, and every uncoupled control and the fault-free trajectory sit at exactly 0.0 pp. Routing costs tens of microseconds per frame. We then asked whether the coupling the method exploits arises on its own. Across eight real RADIATE adverse-weather sequences and three workload proxies independent of the fault signal, after Benjamini-Hochberg correction and a replication run, none of 24 tests found it. We report that null and scope the routing result as a proof of mechanism. Whether such coupling occurs in the field is still open, and the released evaluation pipeline lets a deployment settle it on its own traces.

Sparsh Roy, Vihan Aggarwal, Davin Yin

Massachusetts Institute of Technology, Cambridge, MA, USA · Hopewell Valley Central High School, New Jersey, USA · Millburn High School, New Jersey, USA +1

Adaptive Inference Robotic Perception

Ask PDF

Jul 19, 2026

TAPAS: Throughput-adaptive Perception for Autonomous Systems

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 scene complexity awareness to estimate an appropriate FPS target and 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 KITTI dataset's test sequences, TAPAS achieves 93-100% throughput met rate while saving energy by 76%. On the unseen nuScenes dataset, TAPAS maintains 97% throughput met rate with 64% lower energy compared to SOTA approaches, proving its robustness.

Aman Vyas, Vasista Kodumagulla, Zain Taufique +2

Department of Computing, University of Turku, Finland

Energy-Efficient ML Edge Computing Autonomous Driving Perception Adaptive Computation

Ask PDF

Oct 10, 2026

Towards Privacy-Preserving Thermal Human Perception from Dataset to Deployment

Indoor human perception is the foundation for smart homes and healthcare. However, RGB cameras raise privacy concerns in private indoor environments, while low-resolution thermal cameras provide limited spatial and semantic information. Large-scale indoor thermal datasets and real-world deployments remain limited. Here, we introduce Indoor RGB-T, a large-scale paired RGB--thermal dataset for indoor human perception. The dataset covers 20 indoor rooms and contains 90,576 80x62 thermal frames, 87,016 temporally matched RGB--thermal pairs, and 115,557 annotated human bounding boxes. We establish benchmarks for robust thermal occupant-state detection with 0.75 mAP and occupancy estimation with an accuracy of 0.847 in unseen rooms. We further evaluate RGB-to-thermal and thermal-to-RGB translation, achieving a temperature MAE of 0.654 C and an R2 of 0.923. Finally, we deploy the thermal perception model on a commercial sleep-monitoring system in the real world. Field testing with 100 participants and deployment of approximately 1,000 devices further demonstrate its practical feasibility. We identify practical challenges that motivate future research directions, including environmental thermal interference, sensing range and multi-person occlusion. The code, trained models and partial dataset are available at https://github.com/kailaisun/RGBT and https://huggingface.co/skl24/RGBT.

Kailai Sun, Yue Cui, Xinwei Wang +8

Urban Mobility Lab, Massachusetts Institute of Technology, Cambridge, MA 02139, United States · Singapore-MIT Alliance for Research and Technology Centre (SMART), Massachusetts Institute of Technology, Singapore 138602 · Chongqing General Hospital, Chongqing University, Chongqing 401147, China +4

Thermal Imaging Infrared Object Detection Image-to-Image Translation

Ask PDF

Abstract

Power-Efficient Inference

Explore similar work

Belief-Space Perception Routing under Coupled Sensor Faults and Compute Contention

TAPAS: Throughput-adaptive Perception for Autonomous Systems

Towards Privacy-Preserving Thermal Human Perception from Dataset to Deployment