ActTune: Action-Aware Precision and GPU Operating-Point Adaptation for Energy-Efficient Vision-Language-Action Inference

1 INTRODUCTION

• Reducing inference energy can extend battery-powered robot operation, but lower GPU frequencies prolong inference and can delay action updates.Precision and GPU operating-point control are coupled because precision changes the workload on which operating-point control acts.Source: §1 · p. 1
• A lightweight decision tree selects layer-wise weight and activation precision before each action chunk, while a shared resident weight bank enables switching with one policy inference call per chunk.The tree learns splits and leaf configurations directly from configuration action errors against a BF16 reference policy.Source: §1 · p. 2
• ActTune profiles candidate frequency–power-cap settings under a 10% mean inference-latency-increase limit, then forecasts the next workload and applies the selected setting asynchronously.The comparison is against the same precision selector at the fixed reference GPU operating point.Source: §1 · p. 2
• Quantization sensitivity varies across action classes, Transformer blocks, and weight versus activation precision, while favorable GPU operating points depend on numerical precision.These observations motivate precision selection based on action sensitivity followed by operating-point selection for the resulting workload.Source: §1 · p. 2
• ActTune improves mean task success by up to 2.3% relative to QVLA and ActQuant, and achieves up to 2.02× inference speedup and 76.8% lower energy per successful task versus original BF16 implementations.The reported energy reduction uses GPU operating-point adaptation.Source: §1 · pp. 2–3

2 MOTIVATION

• At fixed A4, C1 achieved 42.5–45.0% task success across tested weight precisions, while C2 maintained 92.5–95.0%; for C1, raising layer-0 activation precision from A4 to A8 raised success to 97.5%.The matched BF16 reference achieved 95.0% mean task success; action classes describe data patterns, not prescribed task stages.Source: §2 · p. 3
• Within the exploratory policy sweep, operating points with at most 1% higher total benchmark time reduced GPU energy by 20.4% for BF16, 22.7% for W4A8, and 21.0% for W4A4.The lowest-energy points used different frequency–power-cap combinations across precision settings.Source: §2 · p. 3, Figure 3 · p. 4
• The hardware sweep used four matched episodes and one measurement per setting, so it provides design evidence rather than a precise estimate of task-level savings.The action clusters are diagnostics; ActTune learns decision regions from configuration action errors.Source: §2 · p. 4

3.1 OVERVIEW

• ActTune selects precision before each policy call and uses GPU profiling to choose operating points under a 10% mean-latency-increase budget.The controller forecasts the next region and configuration, then applies the selected frequency and power cap asynchronously between calls.Source: §3.1 · p. 4, §3.4 · p. 6

3.2 LEARNING DECISION REGIONS FOR PRECISION SELECTION

• The decision tree partitions current proprioception and preceding-action features into regions, each selecting a precision configuration without another policy call.Its regions are learned from configuration action errors rather than predefined semantic action stages.Source: §3.2 · p. 4

3.3 ACTION-CONDITIONED PRECISION WITH SHARED RESIDENT WEIGHTS

• Shared resident weight versions support switchable configurations, while calibration allocates precision under weight, activation, and resident-storage budgets.Sensitivity-guided calibration prioritizes matrices whose quantization affects action predictions and selects tree complexity using validation loss and routing cost.Source: §3.3 · pp. 4–5

3.4 SELECTING GPU FREQUENCY AND POWER CAP

• Because precision changes computation and memory traffic, ActTune profiles frequency and power cap jointly and selects the lowest-energy lookup table meeting its latency budget.The selected table is calibrated on replayed workloads, then frozen and checked on disjoint trajectories.Source: §3.4 · p. 6

3.5 ONLINE EXECUTION AND INTEGER KERNELS

• Online execution keeps settings fixed during inference and runs low-bit linear layers through INT4 or INT8 kernels without reconstructing dense weights.Kernel and runtime optimizations shorten execution time and contribute to energy savings.Source: §3.5 · pp. 6–7

4.1 EXPERIMENTAL SETUP

• The evaluation covers quantized OpenVLA-OFT and π0.5 on four LIBERO suites using NVIDIA RTX A6000 GPUs, reporting success, speed, storage, and energy per successful episode.For each model–budget pair, one tree is shared across the four suites; the π0.5 checkpoint and calibrated bank are also shared.Source: §4.1 · p. 7

4.2 RESULTS

• At W4A4, ActTune reaches 95.5% success on OpenVLA-OFT and 98.4% on π0.5, exceeding the strongest compared baseline by 4.0 and 5.1 percentage points.Its reported storage is 4.47 GB and 3.42 GB, respectively, including all resident weight versions.Source: §4.2 · p. 8
• At W4A16, mean success is 97.0% on OpenVLA-OFT and 98.1% on π0.5, exceeding QVLA and ActQuant on both models.On π0.5, storage is 3.42 GB, 44.8–45.7% below those baselines; energy per successful episode remains above BF16 on both models.Source: §4.2 · p. 8
• Runtime optimization makes π0.5 W4A8 and W4A4 2.02× and 1.81× faster than BF16, while OpenVLA-OFT reaches 1.50× and 1.53×.The largest reported gains are on π0.5, whose quantized implementations were slower than BF16 before optimization.Source: §4.2 · p. 8, Figure 5 · p. 8
• 76.8% lower energy per successful episode on π0.5 W4A8 with GPU operating-point adaptation, with latency increases of 7.79–8.63% across all four settings.All four model–precision settings remain within the 10% inference-latency budget relative to the fixed reference operating point.Source: §4.2 · pp. 7–8, Figure 5 · p. 8

4.3 ABLATION STUDY

• On 400 episodes, GPU operating-point selection saves 26.40% energy, exceeding the best fixed setting and configuration-only control by 4.15 and 4.91 percentage points.Its inference latency increase is 7.79%, compared with 7.93% and 10.00% for those alternatives.Source: §4.3 · p. 8, Figure 6 · p. 9
• For π0.5 W4A4, quantization alone increases energy per success by 27.73%; kernel/runtime optimization and DVFS offset this increase, yielding a net reduction of 56.92%.The ablation measures sequential decreases relative to BF16 energy per success.Source: §4.3 · p. 8, Figure 6 · p. 9
• Both models perform best at K = 3: 97.0% for OpenVLA-OFT W4A8 and 98.4% for π0.5 W4A4, exceeding the single-leaf selector by 1.4 and 1.2 percentage points.K = 4 and K = 6 score 0.6 and 0.4 points below K = 3 for both models.Source: §4.3 · p. 9, Figure 6 · p. 9

• VLA quantization work uses action sensitivity or task evidence for precision allocation, while other methods adapt precision or quantized representations to motion during execution.Related approaches also improve decoding and action chunking, reuse visual tokens or intermediate features, or overlap inference with action execution.Source: §5 · p. 9
• Energy-aware deployment work includes hardware-feedback quantization, combined quantization and DVFS, and GPU frequency adjustment under latency objectives; ActTune evaluates energy per successful episode within a 10% mean-latency-increase constraint.The paper states that its measure includes failed attempts and idle intervals.Source: §5 · p. 9

• Closed-loop LIBERO evaluation reaches 97.0% mean success for OpenVLA-OFT W4A8 and 98.4% for π0.5 W4A4; GPU adaptation reduces energy per successful task by up to 76.8% versus original BF16 implementations.The inference-latency increase remains below 10% relative to the same quantized policies at the fixed reference operating point.Source: §6 · p. 9

• Generative AI tools assisted with scripts for analysis and visualization, literature searches, experimental-result presentation, cross-reference checks, and language polishing.The authors state that they take responsibility for the final code, results, analyses, and manuscript.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The evaluation uses the LIBERO simulation benchmark, not physical-robot experiments.Physical deployment requires task-specific safety assessment and validation beyond the reported simulation experiments.Source: ETHICS STATEMENT · p. 10
• Reducing inference energy may lower robotic-policy resource costs, but does not by itself establish safe behavior.The passage distinguishes energy efficiency from evidence of safety.Source: ETHICS STATEMENT · p. 10
• Physical-robot deployment requires task-specific safety assessment and validation beyond these experiments.Source: ETHICS STATEMENT · p. 10

A.1 EVALUATION SETTINGS AND RESULT PROVENANCE

• 2,000 episodes per model–precision setting across four LIBERO suites underpin the reported closed-loop results, with episodes counted once and suites equally weighted.The decision tree is frozen before evaluation; the table’s comparisons use local evaluations and same deployed checkpoints and episode identities.Source: §A.1 · p. 13

A.2 QUANTIZATION AND EXECUTION SCOPE

• W4A8 and W4A4 are average precision budgets, while eligible layers may use different bit widths and several components retain existing precision.The evaluated π0.5 W4A4 implementation quantizes activations in groups of 64 channels; retained alternatives count toward storage.Source: §A.2 · p. 13, §A.3 · p. 14, Table 2 · p. 14

A.3 RETAINED WEIGHT/ACTIVATION PAIRS

• The evaluated banks retain different subsets of candidate weight/activation pairs; their counts are distinct bit-width pairs, not whole-model configurations.For W4A4, OpenVLA-OFT retains seven pairs and π0.5 retains six; for W4A8, both retain all three candidates.Source: §A.3 · p. 14, Table 2 · p. 14

A.4 INFERENCE TIMING AND MODEL STORAGE

• Inference latency times complete prediction calls on an RTX A6000, while model storage counts persistent compressed tensor payload and resident weight versions.Timing excludes model loading, compilation, graph capture, RPC overhead, and simulation; persistent storage excludes activation buffers and temporary workspaces.Source: §A.4 · pp. 14–15

A.5 ENERGY MEASUREMENT

• Energy per success divides total GPU-board energy for all 2,000 attempted episodes by successes in those same runs.Energy integrates GPU-board power during simulated task execution, including rendering and intervals between calls, but excludes CPU energy and does not measure whole-robot energy.Source: §A.5 · p. 15

A.6 GPU OPERATING-POINT ABLATION COHORT

• 400 episodes per setting, 100 from each LIBERO suite, form the GPU operating-point ablation cohort; the three-leaf precision selector remains fixed.This subset is distinct from the full 2,000-episode success evaluations.Source: §A.6 · p. 15

A.7 CLOSED-LOOP LEAF-COUNT ABLATION PROTOCOL

• K ∈ {1, 2, 3, 4, 6} denotes the tested number of decision-tree leaves, with splits and leaf configurations learned from four configuration losses.One tree shares splits and configuration indices across suites; missing history or invalid features trigger the fallback configuration.Source: §A.7 · p. 15
• Each leaf-count setting uses the same 500 task–initial-state pairs per suite, totaling 10,000 episodes per model across all five settings.The tree is frozen before evaluation, and mean success is reported over the four suites.Source: §A.7 · p. 15

B.1 TRAINING DATA AND HELD-OUT DIAGNOSTICS

• 512 distinct LIBERO training trajectories supply calibration data, with separate diagnostic trajectories and observations excluded from calibration.Sampling uses a frozen selection; evaluation rollouts are not used to fit quantization parameters or selectors.Source: §B.1 · p. 16

B.2 CALIBRATION TARGETS AND SEPARATION FROM EVALUATION

• Calibration uses selected observations to measure quantization error and compare candidate configurations with BF16 actions, while held-out diagnostics and benchmark outcomes remain outside fitting.Baseline adaptations retain their own calibration objectives, and layer coverage, quantizers, and execution backends may differ.Source: §B.2 · p. 16

B.3 PRECISION AND STORAGE BUDGETS

• W4A8 and W4A4 impose separate average active-precision and resident-bank storage constraints rather than uniform layer assignments.For W4A4, the activation bound also applies within each group of layers with the same invocation shape; several components retain existing precision.Source: §B.3 · pp. 16–17

B.4 SELECTING PRECISION FOR EACH LAYER

• The layer allocator ranks low-bit reductions and feasible precision upgrades using reconstruction-error changes, subject to bit and resident-storage budgets.The error proxy ranks assignments rather than establishing globally optimal layer assignments; the deployed banks include model-specific refinements and four frozen configurations.Source: §B.4 · p. 17

B.5 DECISION-TREE INPUTS AND DIAGNOSTIC GROUPS

• The decision tree uses 13 motion and gripper features from the preceding action plus current proprioception, while diagnostic action classes are separate from learned decision regions.The C0–C2 groups organize motivation diagnostics and do not provide labels for the trees.Source: §B.5 · pp. 17–18, Table 4 · p. 18

B.6 ADAPTIVE TREE FITTING AND TRAJECTORY-LEVEL VALIDATION

• Tree fitting selects configuration choices and feature thresholds from weighted configuration losses, then prunes to candidate leaf counts under minimum-child and depth constraints.Validation compares loss and routing cost across candidate sizes without changing fitted splits or leaf choices; the selected tree and bank remain frozen for profiling and evaluation.Source: §B.6 · p. 18

C.1 LEAF-COUNT ABLATIONS ACROSS MODELS

• Three leaves minimize observed validation loss for OFT W4A8, but minima differ across other model–precision settings; the common choice trades some loss for lower routing cost.For OFT W4A8, three leaves reduce validation action loss by 0.182% versus one leaf, with 0.182 µs traversal; three leaves do not minimize loss in every setting.Source: Figure 7 · p. 19, §C.1 · p. 19

C.2 INFERENCE LATENCY BY SUITE

• Table 5 reports per-suite mean inference latency and speedup relative to the original BF16 runtime.The table supplies the per-suite timings underlying four ActTune speedups, which include runtime optimization and do not isolate quantization alone.Source: §C.2 · p. 19, Table 5 · p. 20

C.3 EFFECT OF INFERENCE OPTIMIZATION

• Runtime-optimization comparisons combine CUDA Graph replay, launch-geometry tuning, and model-specific changes, so they do not isolate GEMM kernels or test new closed-loop success.Matching actions on recorded inputs is not a new closed-loop success-rate evaluation.Source: §C.3 · p. 21, Table 6 · p. 21

C.4 ENERGY-SAVING CONTRIBUTION BREAKDOWN

• For π0.5 W4A4, quantization contributes −27.73 percentage points, kernel/runtime optimization +73.88, and DVFS +10.77, totaling 56.92% energy savings.Each component is normalized by the corresponding measured BF16 energy per success; lower power can still coincide with higher energy if execution takes sufficiently longer.Source: Table 7 · p. 21, §C.4 · p. 21

C.5 COMPLETE ACTION-PRECISION INTERVENTION GRID

• The intervention grid quantizes seven language-block projections while leaving vision modules and the action head in BF16; its small, fixed-quantizer cohort does not establish general precision or class-sensitivity rankings.Branches from one task share a reference episode and are not independent trials.Source: §C.5 · pp. 21–22, Table 8 · p. 22

C.6 ACTION-GROUPING DIAGNOSTICS

• Grouping diagnostics report mean silhouette 0.1541 (7.26% negative) and refit adjusted Rand index 0.8455 (range 0.7198–0.9829); these are descriptive, not task-success confidence intervals.All refits retain nine components.Source: §C.6 · p. 22