XS-VLA: Teaching Tiny Vision-Language-Action Models with Spatial Supervision and Demonstration Conditioning

Iok Tong Lei, Ying Jie Yap, Wei Huang, Qingchen Xie, Qianzhi Li, Yujie Zhang, Xiaolong Liu, Zhidong Deng

Compact deployment. Low memory. Effective robot control.

Deployed parameters: 243.99M
Peak inference VRAM: 0.51 GiB
LIBERO success: 90.25%

VRAM measured at batch size 1 in BF16, with two 256 × 256 images and 10 flow steps. On three Mobile ALOHA tasks, average strict success reaches 65.0%, versus 21.7% for the retrained baseline.

XS-VLA on simulated tasks and real robots, including Mobile ALOHA, OpenARM, PiPER, and XLeRobot. 2 min 58 sec, without audio. Download video

We study how spatial supervision and robot demonstrations can help a small vision-language-action policy learn better control. XS-VLA uses these signals during training, then runs with 243.99M parameters and no teacher or posterior encoder.

On LIBERO, the full model reaches 90.25% success in each of two training seeds, compared with 86.00% for our retrained base. On three Mobile ALOHA tasks, average strict success increases from 21.7% to 65.0%.

How XS-VLA is trained

We first train the backbone with coarse spatial labels, then train the policy on robot demonstrations. The two stages provide additional supervision without keeping the teacher or posterior encoder in the deployed model.

XS-VLA architecture: spatial pretraining initializes the vision-language backbone; a posterior encoder maps state and demonstration actions to latent z, which conditions the action flow expert during training. At deployment, z is zero and the posterior encoder is bypassed. — The backbone supplies observation context to the action-space flow expert. A demonstration posterior provides the latent during training; inference uses the prior mean, z = 0.

Coarse-Grained Spatial Distillation

A teacher predicts an object’s location in an image. We map that location to one of nine regions and use the resulting labels to train the compact backbone. The learned parameters then initialize the control policy. We use SmolVLM2-0.25B as a shorthand for the SmolVLM2-256M-Video-Instruct backbone and SmolVLA-0.25B base for our baseline built on it. Here, 0.25B denotes the nominal backbone scale; the complete deployed XS-VLA policy contains 243.99M parameters, including the action expert.

Latent Flow Matching

During policy training, an encoder summarizes the robot state and demonstrated actions into a latent that conditions the flow-matching velocity field. We regularize this posterior toward a Gaussian prior. At inference, we remove the encoder and use the prior mean, z = 0.

More detail on spatial pretraining

CSD pipeline: a teacher predicts object coordinates and confidence; accepted coordinates are quantized into a three-by-three grid; an auxiliary region-label task trains the student backbone for transfer to XS-VLA.

The teacher receives unmarked images. The grid overlay illustrates quantization. Diagnostics evaluate the 14k CSD checkpoint before downstream policy training. They do not establish reliable instruction-conditioned localization; the measured benefit of CSD is in downstream control.

At deployment, the policy takes camera images, an instruction, and the robot state, and produces an action chunk.

LIBERO experiments

We compare against a SmolVLA-0.25B base retrained under our settings, using 160k policy-training steps and a common evaluation protocol. Both the retrained baseline and XS-VLA unfreeze the VLM during policy training.

XS-VLA improves average success by 4.25 percentage points over the retrained base, with the largest gain on the Long suite.

Policy	Spatial	Object	Goal	Long	Overall
SmolVLA-0.25B base	93.0%	88.0%	83.0%	80.0%	86.0%
XS-VLA (original run)	93.0%	91.0%	88.0%	89.0%	90.3%

Reporting precision. As in the paper, LIBERO table success rates are rounded to one decimal place; the text retains more precise values where reported. For example, 361/400 = 90.25%, displayed as 90.3% in the tables. These are the same result. Average episode steps retain two decimal places.

Each run uses 100 evaluation episodes per suite (10 per task), 400 total, with the same evaluation seed and protocol. Training seed 1004 scores 93/93/85/90% on Spatial/Object/Goal/Long and 90.25% overall. Both XS-VLA training seeds achieve 361/400 successes; the retrained base achieves 344/400. Two-seed suite means are 93.0 / 92.0 / 86.5 / 89.5%. Baseline and ablation rows each represent one training run.

Ablations and loss controls

Configuration	Spatial	Object	Goal	Long	Overall	Avg. steps
SmolVLA-0.25B base	93.0	88.0	83.0	80.0	86.0%	180.65
CSD only	82.0	99.0	91.0	83.0	88.8%	177.90
LFM only	93.0	90.0	85.0	82.0	87.5%	177.51
Image-only + LFM	90.0	94.0	82.0	89.0	88.8%	173.85
CSD + Huber, no latent	90.0	89.0	82.0	85.0	86.5%	177.73
XS-VLA (MSE)	94.0	93.0	88.0	84.0	89.8%	173.82
XS-VLA (Huber)	93.0	91.0	88.0	89.0	90.3%	172.54

Suite values are percentages. LFM includes posterior conditioning, KL regularization, and Huber regression. The base and CSD-only configurations use MSE. XS-VLA (MSE) retains the latent branch and KL schedule, changing only the regression loss. Average steps count environment steps per episode, including the effects of success and termination.

The matched image-only pretraining control reaches 88.75% with LFM. Using CSD initialization raises this to 90.25%, providing evidence for a benefit beyond image exposure alone in the evaluated runs.

With CSD and Huber fixed, adding the posterior-conditioning and KL branch raises success from 86.50% to 90.25%. This gain supports the regularized latent branch as a whole; changing the regression loss alone does not explain it.

Experiments on Mobile ALOHA

On three Mobile ALOHA tasks, average strict success rises from 21.7% to 65.0%. Each task uses 100 demonstrations collected by three operators. Each model is evaluated for 20 trials per task.

Task	SmolVLA-0.25B base	XS-VLA
Cup placing	45.0% (9/20)	90.0% (18/20)
95% Wilson interval	[25.8, 65.8]%	[69.9, 97.2]%
Dual-arm block stacking	0.0% (0/20)	30.0% (6/20)
95% Wilson interval	[0.0, 16.1]%	[14.5, 51.9]%
Stacking partial-credit score	3/40	20/40
Sequential coordination	20.0% (4/20)	75.0% (15/20)
95% Wilson interval	[8.1, 41.6]%	[53.1, 88.8]%
Average strict success	21.7%	65.0%

Both models use 80k policy-training steps. XS-VLA starts from the same 14k CSD checkpoint used for LIBERO. Strict stacking success requires a stable completed stack. Partial credit awards one point for bringing two blocks together and two points total for a completed stable stack (40 points maximum). The average uses only the three strict success rates. Wilson intervals assume independent Bernoulli trials.

The demonstration video also shows XS-VLA operating on OpenARM, PiPER, and XLeRobot. These platform-specific deployments provide qualitative evidence across robot platforms.

Watch the video

Deployment footprint and inference cost

The full checkpoint has 244.83M parameters. Removing 0.84M training-only parameters leaves a 243.99M-parameter deployed policy, including the action expert.

GPU	Peak VRAM	Latency / chunk	Query rate
A100 PCIe 40GB	0.51 GiB	259.93 ms	3.85 Hz
RTX 3090	0.51 GiB	321.08 ms	3.11 Hz

Peak inference VRAM: 0.51 GiB on both GPUs. Batch size 1; two 256 × 256 images; BF16; 10 flow steps; 8 warm-up runs and 50 measurements. Tokenization, simulation, and video writing are excluded.

LIBERO generates 50 actions per query and executes 10 before replanning. Query rate measures model inference; it is distinct from amortized action throughput and robot servo frequency.

The paper includes the full training setup, spatial diagnostics, and limitations.

Citation

@misc{lei2026xsvlateachingtinyvisionlanguageaction,
      title={XS-VLA: Teaching Tiny Vision-Language-Action Models with Spatial Supervision and Demonstration Conditioning}, 
      author={Iok Tong Lei and Ying Jie Yap and Wei Huang and Qingchen Xie and Qianzhi Li and Yujie Zhang and Xiaolong Liu and Zhidong Deng},
      year={2026},
      eprint={2607.04171},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.04171}, 
}