Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning

Seongmin Park; Jaewon Lee; Yujun Heo; Yoonsik Kim; Seheon Choi; Hyoung-rock Kim; Jungwook Choi

cs.ROOct 11, 2026

Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning

Authors: Seongmin Park, Jaewon Lee, Yujun Heo, Yoonsik Kim, Seheon Choi, Hyoung-rock Kim, Jungwook Choi

Organizations: Hanyang University, Republic of Korea · LG Electronics, Republic of Korea

Abstract

Quantization lowers deployment costs of vision-language-action (VLA) policies but can compromise task completion. Across five VLA models and 4,050 real-world trials, we systematically reveal that quantization-aware training (QAT) generally improves on post-training quantization (PTQ) but leaves substantial gaps to full precision (FP) on long-horizon tasks and under visual and semantic shifts, while its imitation-based objective does not explicitly use rollout success or partial completion. To exploit these outcome signals, we introduce Quantization-Conditioned Reinforcement Learning (QCRL), an offline RL method coupling progress-based advantage estimation with quantization-aware source conditioning. FP rollouts supply reference experience beyond action-level distillation, progress-scaled terminal penalty distinguishes partial completion, and quantization-aware conditioning selects advantageous reference behavior within the quantized policy. We evaluate weight-only INT4 W4A16 on $π_{0.5}$ and GR00T N1.5 and joint weight-activation NVFP4 W4A4 on $π_{0.5}$ . Using only in-distribution data, QCRL reaches 99.3% of FP's mean score across three visual-shift conditions for $π_{0.5}$ with INT4 W4A16, alongside 42.9% model-memory savings. With NVFP4 W4A4, it reaches 100.7% of the same FP baseline, alongside a $1.73\times$ speedup using NVFP4/FP8 deployment. Hardware gains are relative to FP16 on Jetson Thor.

Figures & tables

Fig. 1: QCRL components: (a) FP rollout augmentation; (b) progress-scaled terminal penalty and advantage labels; (c) quantization-aware conditioning and policy learning. Dataset actions provide targets for $\mathcal{L}_{QCRL}$ (Eq. (8)); deployment selects the positive reference condition.

Fig. 2: Real-world task sequences (left to right): short tasks, folding, and clean desk in rows 1–3. Bottom row: visual and semantic OOD conditions.

Setting	$\pi_{0}$ , $\pi_{0.5}$	$\pi_{0}$ -FAST	SmolVLA	GR00T N1.5
Chunk/batch	50 / 2	32 / 2	50 / 64	16 / 32
FP	100k / 2.5	140k / 5	3.2k / 20	16k / 10
QAT	50k / 2	70k / 2	1.6k / 20	8k / 10

TABLE I: FP/QAT training: steps / learning rate (

10^{-5}

). openpi (JAX):

\pi

models [ 23 ] ; LeRobot: SmolVLA/GR00T [ 24 ] .

Condition	Metric	$\pi_{0}$			$\pi_{0.5}$			$\pi_{0}$ -FAST			SmolVLA			GR00T N1.5
		FP	PTQ	QAT	FP	PTQ	QAT	FP	PTQ	QAT	FP	PTQ	QAT	FP	PTQ	QAT
Sweep	Full	$96.7$	$100.0$	$100.0$	$100.0$	$100.0$	$100.0$	$83.3$	$80.0$	$83.3$	$93.3$	$70.0$	$83.3$	$100.0$	$93.3$	$96.7$
Pick/place	Full	$83.3$	$73.3$	$86.7$	$76.7$	$63.3$	$76.7$	$40.0$	$26.7$	$36.7$	$30.0$	$16.7$	$23.3$	$50.0$	$30.0$	$40.0$
Stack	Full	$63.3$	$40.0$	$63.3$	$83.3$	$46.7$	$83.3$	$26.7$	$10.0$	$23.3$	$23.3$	$3.3$	$13.3$	$20.0$	$10.0$	$13.3$
Folding	Full	$70.0$	$10.0$	$60.0$	$40.0$	$0.0$	$0.0$	$0.0$	$0.0$	$0.0$	$0.0$	$0.0$	$0.0$	$40.0$	$30.0$	$30.0$
Folding	Progress	$80.0$	$25.0$	$80.0$	$60.0$	$15.0$	$40.0$	$15.0$	$0.0$	$5.0$	$10.0$	$0.0$	$0.0$	$55.0$	$40.0$	$45.0$

TABLE II: Real-world task performance and OOD generalization across five VLA models (%). PTQ and QAT use INT4 W4A16.

Fig. 3: Quantization effects across five VLA models. (a) Mean full success rate; long horizon averages folding and clean desk. (b) ID averages three tasks; visual averages 11 task–condition scores. Each mean is normalized to the model’s corresponding FP mean (

\mathrm{FP}=1

), retaining ratios above one. (c) Primitive progress scores for object and recipient replacement; markers identify models, and dashed lines average all five models across both shifts.

(a) Mean recorded-action $L_{2}$ error (action source / $q_{t}$ )
Policy	Demo / $\mathtt{ref}$	Demo / $\mathtt{quant}$	FP / $\mathtt{ref}$	FP / $\mathtt{quant}$
PTQ	$1.852\pm 0.731$	$1.852\pm 0.731$	$1.969\pm 1.056$	$1.969\pm 1.056$
QCRL	$1.265\pm 0.675$	$1.305\pm 0.658$	$1.456\pm 0.878$	$1.495\pm 0.913$

TABLE III: QCRL diagnostics. (a):

\pi_{0.5}

W4A16, mean

\pm

episode SD. PTQ uses no

q_{t}

. (c): mean

\pm

observation SD.

Item	Setting
Real-world data	Per task: short 150/50/50; long 80/30/30
Task grouping	One policy for three short tasks; one each for folding and clean desk
LIBERO data	Per model: four suites (40 tasks), 1,693/2,000/2,000 (5,693 total)
	FP/PTQ: 500 per suite each; seeds 42/7
Evaluation trials	Real: short ID 30, OOD/long 10; LIBERO: 500 per task
Episode limits	LIBERO Spatial/Object/Goal: 480; Long: 520 steps

TABLE IV: QCRL data, evaluation, and training settings. Collected episodes: Demo/FP/PTQ; batch sizes: global/micro.

Task	Instruction
Sweep	sweep the gray cloth to the right side of the table
Pick/place	pick up the corn and place it on the plate
Stack	pick up the pink cup and stack it inside the blue cup
Folding	fold the blue cloth in half twice
Clean desk	clean the desk by putting items into the green basket and trash into the yellow box

TABLE V: Real-world ID task instructions.

Fig. 4: Cloth folding ( $\pi_{0.5}$ , INT4 W4A16): QAT fails at the second fold; QCRL completes it. Additional tasks are shown in the supplementary video.

Task	$\pi_{0.5}$				GR00T N1.5
Task	FP	PTQ	QAT	QCRL	FP	PTQ	QAT	QCRL
Sweep	$100.0$	$100.0$	$100.0$	$100.0$	$100.0$	$93.3$	$96.7$	$100.0$
Pick/place	$76.7$	$63.3$	$76.7$	$80.0$	$50.0$	$30.0$	$40.0$	$40.0$
Stack	$83.3$	$46.7$	$83.3$	$83.3$	$20.0$	$10.0$	$13.3$	$23.3$
Folding (full)	$40.0$	$0.0$	$0.0$	$30.0$	$40.0$	$30.0$	$30.0$	$40.0$
Folding (prog.)	$60.0$	$15.0$	$40.0$	$60.0$	$55.0$	$40.0$	$45.0$	$60.0$

TABLE VI: Real-world ID performance (%; quantized: INT4 W4A16). Full success unless marked prog. (progress)

Condition	$\pi_{0.5}$				GR00T N1.5
Condition	FP	PTQ	QAT	QCRL	FP	PTQ	QAT	QCRL
Appearance	$86.7$	$73.3$	$83.3$	$86.7$	$51.7$	$41.7$	$50.0$	$55.0$
Illumination	$73.3$	$46.7$	$53.3$	$73.3$	$63.3$	$55.0$	$55.0$	$86.7$
Sensor degradation	$70.0$	$13.3$	$41.7$	$68.3$	$45.0$	$21.7$	$20.0$	$45.0$
Unseen object	$87.5$	$45.0$	$67.5$	$85.0$	$52.5$	$42.5$	$45.0$	$67.5$
Role reversal	$43.3$	$43.3$	$40.0$	$60.0$	$33.3$	$33.3$	$30.0$	$40.0$

TABLE VII: OOD and hardware results (quantized: INT4 W4A16). ID-only adaptation. Visual/semantic scores (%): task score/primitive progress.

Type	Condition	FP		NVFP4 W4A4
Type	Condition	Base	Adapt.	PTQ	QAT	QCRL
Simulation	LIBERO-Long	$93.2$	$94.1$	$92.5$	$93.0$	$94.0$
Visual	Appearance	$86.7$	$86.7$	$48.3$	$60.0$	$83.3$
	Illumination	$73.3$	$70.0$	$40.0$	$70.0$	$73.3$
	Sensor degradation	$70.0$	$73.3$	$43.3$	$70.0$	$75.0$
Semantic	Unseen object	$87.5$	$87.5$	$75.0$	$80.0$	$80.0$

TABLE VIII: Task and hardware results for

\pi_{0.5}

NVFP4 W4A4. Scores (%): simulation success, visual task score, semantic primitive progress. FP Base/Adapt.: task-trained baseline/FP QCRL control.

Method	$\pi_{0.5}$				GR00T N1.5
Method	Spatial	Object	Goal	Long	Spatial	Object	Goal	Long
FP	$99.0$	$98.8$	$98.2$	$93.2$	$96.8$	$98.4$	$95.4$	$87.0$
FP RECAP	$99.8$	$99.0$	$98.6$	$94.4$	$96.8$	$98.2$	$95.6$	$87.4$
PTQ	$98.8$	$98.4$	$97.8$	$92.4$	$96.2$	$97.4$	$92.0$	$85.2$
RECAP+QAT	$99.2$	$99.0$	$98.5$	$93.0$	$96.2$	$97.6$	$92.2$	$85.8$
QCRL	$99.4$	$99.2$	$98.4$	$94.8$	$96.8$	$98.6$	$95.8$	$87.8$

TABLE IX: LIBERO full success (%; quantized methods: INT4 W4A16).

FP roll.	–	✓	–	–	✓	✓	–	✓
Prog.	–	–	✓	–	✓	–	✓	✓
Prec. cond.	–	–	–	✓	–	✓	✓	✓
Long	93.0	94.0	93.4	93.2	94.0	94.4	93.8	94.8
Sensor	58.3	65.0	60.0	60.0	66.7	66.7	63.3	68.3

TABLE X: QCRL ablations (

\pi_{0.5}

, INT4 W4A16; %). All-off/on: RECAP+QAT/QCRL. Prec. cond.: quantization-aware conditioning. Long/Sensor: LIBERO-Long success/sensor-degradation task score.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Group	Condition	Change from the ID setting	Interpretation of the reported score
Visual	Appearance	Unseen cloth color or unseen background.	Sweeping success; pick/place or stacking progress.
	Illumination	Reduced image brightness/contrast.	Same task metrics under changed observations.
	Sensor degradation	Blurred observations.	Same task metrics under degraded observations.
	Distractors	Additional task-irrelevant objects.	Pick/place or stacking progress.
Semantic	Object / recipient	Corn $\rightarrow$ carrot; plate $\rightarrow$ pot.	Mean completion of pick and place stages.
	Role reversal	Reverse the pink and blue cups’ instructed roles.	Main-paper primitive progress averages move-to-target, pick, and place, without certifying correct roles.

Appendix

TABLE A1: Additional OOD evaluation definitions. Each evaluated task–condition pair uses 10 trials per model and method. The main paper defines full success and milestone progress.

Task	Condition	Input instruction
Pick-and-place	Added context	before doing anything else, look at the carrot, then pick up the corn and place it on the plate
Pick-and-place	Null input	null
Stacking	Added context	you only need to focus on the cups. pick up the pink cup and stack it inside the blue cup
Stacking	Null input	null

Appendix

TABLE A2: Added-context and null-input prompts used only for semantic evaluation.

Visual condition	Included settings	FP	PTQ	QAT
Appearance shift	Unseen color / unseen background	60.3 (0.0)	48.0 (12.3)	56.3 (4.0)
Illumination shift	Brightness / contrast shift	61.7 (0.0)	47.0 (14.7)	51.7 (10.0)
Sensor degradation	Blurred observation	48.0 (0.0)	19.0 (29.0)	28.7 (19.3)
Distractor	Additional irrelevant objects	16.0 (0.0)	8.5 (7.5)	10.5 (5.5)

Appendix

TABLE A3: Visual generalization task-score summary across the four visual shifts. Scores are averaged over evaluated task and model entries. For cloth sweep the score denotes full-task success, and for pick-and-place and cup stacking it denotes the ordered-milestone task score. All quantized methods use INT4 W4A16. Each cell reports score with drop from FP in parentheses (

\Delta=\mathrm{FP}-\mathrm{Method}

); a smaller parenthetical means closer to FP.

Task	Shift	Method	$\pi_{0}$	$\pi_{0.5}$	$\pi_{0}$ -FAST	SmolVLA	GR00T N1.5
Cloth sweep	Unseen color	FP	100	100	80	80	80
		PTQ	100	100	70	70	60
		QAT	100	100	70	80	80
	Illumination	FP	100	100	70	100	70
		PTQ	100	50	50	90	60
		QAT	100	60	60	100	60

Appendix

TABLE A4: Visual OOD and robustness generalization results over 10 trials. All values are reported as percentages. For cloth sweep, the score denotes full-task success. For pick-and-place and cup stacking, the score denotes the average progress over ordered milestones. All quantized methods use INT4 W4A16.

Task	Metric	FP	PTQ	QAT
Sweep	Full	100.0	86.7	90.0
Pick/place	Full	93.3	80.0	90.0
Stack	Full	46.7	26.7	36.7
Folding	Full	0.0	0.0	0.0
	Progress	30.0	5.0	20.0
Clean desk	Full	0.0	0.0	0.0

Appendix

TABLE A5: Additional longer-trained SmolVLA results (%).

Task	Condition	Method	$\pi_{0}$	$\pi_{0.5}$	$\pi_{0}$ -FAST	SmolVLA	GR00T N1.5
Pick/place	Corn $\rightarrow$ carrot	FP	90.0	90.0	10.0	45.0	60.0
		PTQ	65.0	50.0	5.0	10.0	50.0
		QAT	85.0	75.0	10.0	25.0	50.0
Pick/place	Plate $\rightarrow$ pot	FP	55.0	85.0	5.0	35.0	45.0
		PTQ	45.0	40.0	0.0	10.0	35.0
		QAT	50.0	60.0	0.0	25.0	40.0

Appendix

TABLE A6: Full task-level semantic-shift and null-input diagnostic results (%). Rows retain separate FP, PTQ, and QAT scores. Quantized methods use INT4 W4A16. Null-input values describe task-like progress, not success for a valid instruction.

Task / condition	Score metric	FP		PTQ		QAT
Task / condition	Score metric	Score $\uparrow$	ACE $\downarrow$	Score $\uparrow$	ACE $\downarrow$	Score $\uparrow$	ACE $\downarrow$
Cloth sweeping / ID	Full	100.0	112.2	100.0	116.2	100.0	112.9
Pick-and-place / ID	Full	76.7	115.2	63.3	121.5	76.7	117.2
Cup stacking / ID	Full	83.3	123.9	46.7	131.5	83.3	125.2
Pick-and-place / sensor degradation	Milestone	80.0	189.6	10.0	204.5	55.0	198.3

Appendix

TABLE A7: Action-side diagnostic results for

\pi_{0.5}

. Scores are percentages; “Full” denotes full task success and “Milestone” denotes milestone progress. ACE is reported separately from task score and is not a success-rate estimate.

Task	View	Mask type	Threshold	Prompt set
Pick-and-place	Base	Action	0.18	target-object prompts: small yellow corn, yellow corn kernel, small corn on metal bowl, yellow corn kernels in bowl
Pick-and-place	Wrist	Action	0.16	target-object prompts: yellow corn cob only, yellow plastic corn kernel, small yellow corn on table, yellow corn cob tip only
Pick-and-place	Base	Support	0.26	receptacle prompts: metal bowl, metal plate, bowl rim, metal lip
Pick-and-place	Wrist	Support	0.11	receptacle prompts: metal bowl, metal plate, bowl rim, metal lip
Cup stacking	Base	Action	0.08	manipulated-cup prompts: pink cup, red cup, pink top-view cup disk, red top-view cup disk, target cup
Cup stacking	Wrist	Action	0.02	manipulated-cup prompts: gripper-held cup, held pink cup, pink cup rim, red cup rim, cup in gripper

Appendix

TABLE A8: Task-specific action/support CLIPSeg prompt sets and thresholds used for NMR evaluation.

Fig. 5: Nuisance-mask and perturbation-attribution example used to illustrate the NMR diagnostic. Numerical comparisons are reported in Table A.8 .

TABLE A9: NMR comparisons for $\pi_{0.5}$ . Lower values indicate less saliency mass in designated nuisance regions.

Model	Axis	FP	PTQ	QAT
$\pi_{0}$	ID	81.1	71.1 (10.0)	83.3 ( $-2.2$ )
	Long	73.8	31.3 (42.5)	71.3 (2.5)
	Visual	58.6	45.0 (13.6)	51.4 (7.3)
$\pi_{0.5}$	ID	86.7	70.0 (16.7)	86.7 (0.0)
	Long	65.0	27.5 (37.5)	46.3 (18.8)
	Visual	76.4	44.1 (32.3)	57.3 (19.1)

Appendix

TABLE A10: Derived model-level robustness profiles (%). Parentheses show the drop from FP in percentage points, computed before rounding.

Pair	Matched-noise $L_{2}$	Fresh-noise $L_{2}$
$\pi_{\mathrm{FP}}^{0}\leftrightarrow\pi_{Q}^{0}$	$0.87\pm 0.21$	$1.38\pm 0.46$
$\pi_{\mathrm{FP}}^{0}\leftrightarrow\pi_{\mathrm{FP}}^{0}$	—	$1.06\pm 0.39$
$\pi_{Q}^{0}\leftrightarrow\pi_{Q}^{0}$	—	$1.14\pm 0.40$
$\pi_{\mathrm{ext}}\leftrightarrow\pi_{Q}^{0}$	$18.07\pm 3.57$	$18.37\pm 3.82$
$\pi_{\mathrm{ext}}\leftrightarrow\pi_{\mathrm{FP}}^{0}$	$18.11\pm 3.54$	$18.37\pm 3.78$
$\pi_{\mathrm{ext}}\leftrightarrow\pi_{\mathrm{ext}}$	—	$11.41\pm 4.11$

Appendix

TABLE A11: Complete rollout-source action-distance comparison. Entries are mean

\pm

observation SD over 100 shared PTQ observations. Within-policy rows use independent seeds and characterize sampling variability; they are not guaranteed lower bounds on between-policy distances.

VLM Quantization

Vision-language model quantization reduces model precision to lower memory and computation while preserving multimodal capabilities.

See all 52 papers

VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models
Sep 21, 2026Jiuyi Xu, Qing Jin, Meida Chen +3Mixed-Precision Quantization Closed-Loop Control of VLA Models
Colorado School of Mines · Independent Researcher · University of Central Florida +2
CHASE-VLA: Post-Training Quantization Framework for Vision-Language-Action Models with Chunk-Aware Scale Estimation
Oct 2, 2026Jin Hyun, Jung Gyu Min, Gyuhyun Jung +1Post-Training Quantization VLM Quantization
KAIST, Daejeon, South Korea · POSTECH, Pohang, South Korea
Ω-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling
May 27, 2026Xinyu Wang, Mingze Li, Sicheng Lyu +6Post-Training Quantization Efficient VLA Model Inference
McGill University · Université de Montréal · Mila – Quebec AI Institute +2

Sep 21, 2026J/K move · Enter open · S save

VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models

Jiuyi Xu, Qing Jin, Meida Chen, Song Wang+2

Colorado School of Mines · Independent Researcher · University of Central Florida+2

Mixed-Precision Quantization Closed-Loop Control of VLA Models Efficient VLA Model Inference Vision-Language-Action Models Post-Training Quantization VLA Model Post-Training

Open paper AskPDF arXiv

Post-training quantization reduces the memory requirements of vision-language-action (VLA) models, but precision selection must account for the interaction between layer scope, numerical format, and calibration. We introduce VLAQuantBench, a controlled evaluation with 409 runs and 94,574 simulation episodes: four models on LIBERO, with X-VLA additionally evaluated on three simulation benchmark families. Under uncalibrated W4A4 round-to-nearest quantization, expanding a

π_{0.5}

action-head subset from 126 to 167 layers raises success from 7.0% to 70.5%. Fixed-observation replay confirms a corresponding numerical recovery. Two-episode calibration removes the severe joint failures in the tested subsets, whereas the same smoothing-and-clipping recipe lowers

π_0

success and does not recover OpenVLA-OFT end-to-end. For OpenVLA-OFT, protecting one 28,672-parameter output projection instead restores near-baseline success: the remaining 441 eligible linear layers retain W3 on LIBERO-Long or eight-bit activations across all four suites. Task-clustered intervals support the large failure and recovery contrasts. These results establish recipe-dependent interactions and identify concrete precision assignments, rather than universal layer-sensitivity rules. Real-kernel and physical-robot measurements complement the accuracy analysis. Code, configurations, and episode records are publicly available at https://github.com/jiuyixu25/VLAQuantBench.

Abstract

VLM Quantization

Explore similar work

VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models

CHASE-VLA: Post-Training Quantization Framework for Vision-Language-Action Models with Chunk-Aware Scale Estimation

Ω-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling