CLP-VLA

FIRST_AUTHOR_LAST, FIRST_AUTHOR_FIRST; SECOND_AUTHOR_LAST, SECOND_AUTHOR_FIRST

Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

Gia-Binh Nguyen^1,2,*, Trong-Bao Ho², Thien-Loc Ha², Khoa Vo³, Philip Lund Møller⁴, Quang T. Nguyen², Long Dinh^1,2, Tung M. Luu⁶, Tuan Dam⁵, Vu Duong¹, Trung Le⁷, Nghi D. Q. Bui¹, Minh Vu^1,2, Tran Nguyen Le⁴, An Thai Le^1,2,13, Ngan Le³, Daniel Sonntag^8,9, James Zou¹², Jan Peters^9,13, Duy M. H. Nguyen^†,9,10,11, Ngo Anh Vien^†,1,2

¹Center for AI Research, VinUniversity ²VinRobotics ³University of Arkansas ⁴Technical University of Denmark ⁵Hanoi University of Science and Technology ⁶KAIST ⁷Monash University ⁸Oldenburg University ⁹DFKI ¹⁰University of Stuttgart ¹¹IMPRS-IS ¹²Stanford University ¹³Technische Universität Darmstadt
^†Project Leads.

Paper Coming Soon Code arXiv

Abstract

VLA models are far more compressible than we assumed.

Validated on

GR00T-N1.5

π0

SmolVLA

Inference

Up to 66%

layers pruned

Memory

↓33%

footprint reduced

Training

40–50%

time & cost saved

Despite their remarkable capabilities, Vision-Language-Action (VLA) models impose prohibitive computational costs at fine-tuning and deployment time. We reveal a highly non-trivial architectural property of leading continuous-control foundations (π0, GR00T-N1.5): despite training on diverse physical trajectories, these models exhibit severe layer-wise representational redundancy.

To exploit this, we introduce CLP — a training-free structural compression pipeline. Using a single forward pass via Centered Kernel Alignment (CKA), CLP identifies and removes redundant twin layers, permanently compressing model depth by up to 50% across both the VLM backbone and the continuous-control policy head, without any auxiliary routing or early-exit modules.

Downstream fine-tuning of the streamlined architecture delivers a dual benefit: 40–50% faster training and up to 30% faster real-time inference, while matching or exceeding full-scale performance. We validate across three simulation benchmarks (LIBERO, RoboCasa, SimplerEnv) and 10 real-world manipulation tasks spanning 4 robotic embodiments — proving that advanced VLAs require significantly fewer layers than previously assumed.

Contributions

Three concrete advances in efficient VLA adaptation.

Accessible adaptation of SOTA VLAs

π0 and GR00T-N1.5 fine-tuned at reduced depth — lower memory, training, and inference cost with no auxiliary modules.

Pre-finetuning compression via CLP

CKA identifies redundant transformer blocks and removes them entirely before adaptation — no retraining required.

Multi-embodiment validation

Validated on 3 sim benchmarks and 10 real-world tasks across 4 robots. CLP is platform-agnostic and acts as an effective structural regularizer under limited data.

Method

One forward pass. No extra modules. Up to 50% faster training.

0:00

▮▮

CKA-Guided Layer Pruning (CLP)

CKA Profiling

Feed sample batches through the pretrained VLA and compute pairwise CKA similarity across all layers.

Block Identification

Group consecutive high-similarity layers (CKA ≥ τ) into redundant blocks; keep only the first layer of each.

Training-free Compression

Prune 33–66% of transformer layers with ~30% FLOP reduction — no routing, no auxiliary modules.

Fine-tune & Deploy

Fine-tune the pruned model on the target task. Training time cuts up to 50%; inference up to 30–50% faster.

Key Findings

What we discovered inside large VLA models.

Finding 01

VLA layers are highly redundant

CKA heatmaps reveal large contiguous blocks in π0 and GR00T-N1.5 producing nearly identical representations — spanning the VLM backbone, action expert, and DiT heads.

Finding 02

Up to 2/3 of layers can be pruned with minimal loss

Both π0 and GR00T-N1.5 maintain flat performance up to 50% pruning, achieving ×1.39–1.42 FLOPs reduction with negligible success rate drop.

Finding 03

Trains faster — and often performs better

CLP trains significantly faster while matching or exceeding baseline success rates — e.g. 75% vs 65% on BananaToPot in less time.

Finding 04

Pruned models recover base model behavior after finetuning

Hidden state PCA shows representations realign closely with the base model post-finetuning — confirming CLP preserves representational structure despite removing up to half the layers.

Experiments & Results

Validated across simulation and 4 real-world robot embodiments.

Real-world Manipulation

UR10 — Single Arm

Groceries ToBasket

Open Kettle

Close Kettle

UR5 — Single Arm

Serve Napkin

ScrewdriverToBasket

ALOHA — Single Arm

BananaToPot

CubeToDrawer

Block Stacking

ALOHA — Bimanual

Fold Shorts

Fly Towel

Task success rate (%) and training efficiency across 4 robot embodiments.

Robot	Task	SR (%)	Time (h)	SR-CLP (%)	Time-CLP (h)	Saved
UR10 Single Arm	Groceries ToBasket	90	11.8	89	8.0	32%
	Open Kettle	100	2.8	95	1.4	50%
	Close Kettle	100	3.0	100	1.5	50%
UR5 Single Arm	Serve Napkin	45	1.1	65	0.7	36%
UR5 Single Arm	ScrewdriverToBasket	15	1.5	30	1.1	27%
ALOHA Single Arm	BananaToPot	65	5.1	75	2.9	43%
	CubeToDrawer	75	5.6	60	3.2	43%
	Block Stacking	80	2.8	75	1.4	50%
ALOHA Bimanual	Fold Shorts	90	6.5	95	4.4	32%
ALOHA Bimanual	Fly Towel	75	3.2	70	2.1	34%
Average / Total		73.5	43.4h	75.9	27.7h	36%

Simulation Experiments

LIBERO

RoboCasa

SimplerEnv

LIBERO benchmark — comparison with training-free acceleration methods.

Method	Spatial	Object	Goal	Long	Avg SR	Speedup ↑
OpenVLA-OFT group
FastV	94.6	95.8	94.0	88.8	93.3	1.44×
DivPrune	92.4	91.2	89.0	84.8	89.4	1.46×
EfficientVLA	96.5	91.1	96.0	72.1	88.9	1.52×
π0 group
π0	94.6	98.2	95.4	90.0	94.6	1.00×
π0-SpecPrune-VLA	96.6	98.0	95.2	84.2	93.5	1.31×
π0-CLP Ours	95.0	99.2	95.0	86.4	93.9	1.39×
GR00T-N1.5 group
GR00T-N1.5	90.8	98.4	95.4	91.0	93.9	1.00×
GR00T-N1.5-CLP Ours	89.4	98.8	95.8	88.6	93.0	1.42×
SmolVLA group
SmolVLA	71.8	92.2	87.4	57.2	77.15	1.00×
SmolVLA-CLP Ours	75.6	93.0	81.6	56.2	76.75	1.47×

GR00T-N1.5 on SimplerEnv (WidowX) — task success rate (%).

Task	GR00T-N1.5	GR00T-N1.5-CLP Ours
Carrot Plate	26	34 ↑8%
Eggplant Basket	34	14
Spoon Towel	18	38 ↑20%
Stack Cube	8	4
Eggplant Sink	8	16 ↑8%
Close Drawer	12	24 ↑12%
Open Drawer	10	10
Average	16.57	20.0 ↑3.43%
Training Time (hrs)	22.9	15.7 ↓31%

π0 on LIBERO — trained on 10% data (% Success Rate).

π0 on RoboCasa — 30 demonstrations (% Success Rate).