README.md

metadata

license: apache-2.0
language:
  - en
library_name: lerobot
pipeline_tag: robotics
tags:
  - robotics
  - vla
  - vision-language-action
  - edge-ai
  - manipulation
  - fastvit
  - smolvla
  - lerobot
  - jetson
  - edge-deployment
datasets:
  - lerobot/fmb
base_model:
  - lerobot/smolvla_base
  - HuggingFaceTB/SmolVLM2-500M-Video-Instruct
model-index:
  - name: EdgeVLA-Small
    results:
      - task:
          type: robotics
          name: Action Prediction (FMB)
        dataset:
          type: lerobot/fmb
          name: Functional Manipulation Benchmark
        metrics:
          - type: mse
            value: 0.515
            name: Action MSE
          - type: accuracy
            value: 95.8
            name: Gripper Accuracy (%)
          - type: cosine_similarity
            value: 0.679
            name: Cosine Similarity

EdgeVLA-Small (FMB)

228M parameter edge-optimized Vision-Language-Action model. 49% smaller than SmolVLA, 47% faster, 17% better accuracy. Best efficiency/accuracy tradeoff. Trained on real-robot data.

EdgeVLA-Small combines FastViT-t8 vision with aggressive VLM layer pruning (16 to 8 layers) to achieve a 228M model that beats the 450M SmolVLA baseline on action prediction while running nearly 2x faster. The vision encoder is trained end-to-end — every parameter contributes at inference. Architecture inspired by DynamicVLA; VLM layer pruning is our contribution.

Trained exclusively on lerobot/fmb (3-camera Franka Panda manipulation). Source code: enfuse/edgevla

Intended Use & What You Can Do With This Model

This model predicts 7-DoF robot actions (x, y, z, rx, ry, rz, gripper) from 3 camera images. It outputs 50-step action chunks at 10Hz — each inference produces 5 seconds of continuous robot motion.

Immediate uses:

• Deploy on a Franka Panda (or compatible 7-DoF arm) with a 3-camera setup for FMB-style tabletop manipulation. Feed camera frames in, execute the predicted delta actions.
• Fine-tune on your own robot data — this is the most practical use. If you have any robot with cameras in LeRobot format, this checkpoint is an excellent pretrained starting point. Fine-tuning at LR=3e-5 for 50K steps typically adapts well to new setups.
• Edge deployment — the best efficiency/accuracy tradeoff in the EdgeVLA family. Estimated ~184ms on Jetson Orin AGX with TensorRT FP16. At 228M params and 435MB FP16, it fits comfortably on Jetson Orin NX (16GB) and AGX (32–64GB).
• Research baseline — 49% smaller than SmolVLA, nearly 2x faster, and still 17% better on action prediction.

Important caveats:

• All metrics below are offline action prediction on held-out FMB samples. There are no closed-loop success rate numbers — the model has not been validated on a physical robot completing full tasks.
• Trained specifically on FMB data (Franka Panda, specific manipulation tasks, 3-camera setup). It will not generalize to different robots, camera configurations, or tasks without fine-tuning.
• The model expects 3 camera inputs (side_1, side_2, wrist). For single-camera setups, you would need to fine-tune with --empty_cameras or retrain.

Results (FMB Offline, 500 held-out samples)

Metric	SmolVLA (450M)	EdgeVLA-Small (228M)	Delta
Action MSE	0.618	0.515	-17%
Cosine Similarity	0.663	0.679	+2%
Gripper Accuracy	94.9%	95.8%	+0.9pp
Inference Latency (H200)	169ms	90ms	-47%
Memory (FP16)	858MB	435MB	-49%

Per-Dimension MSE

Dim	SmolVLA	Small	Delta
x	0.538	0.514	-4%
y	0.598	0.540	-10%
z	0.599	0.529	-12%
rx	0.624	0.462	-26%
ry	1.358	0.974	-28%
rz	0.373	0.395	+6%
gripper	0.233	0.192	-18%

Latency (H200, FP32)

Mean	P50	P95	Throughput
90ms	87ms	105ms	11.1 Hz

Architecture

EdgeVLA-Small (228M total, 55M trainable):
  FastViT-t8 vision:        4.0M  (trainable, replaces SigLIP 98M frozen)
  VLM (SmolLM2-360M):    169.2M  (frozen, 8 layers — pruned from 16)
  Action expert:           49.1M  (trainable, flow matching, 0.75x width)
  Projections:              1.6M  (trainable)

Key changes from SmolVLA: FastViT-t8 (conv, trainable) replaces SigLIP (ViT, frozen). VLM pruned 16 to 8 layers. 64 visual tokens vs 729 (11x fewer). 256x256 input vs 384x384.

Training

Parameter	Value
Dataset	lerobot/fmb
Total steps	150K (50K + 100K continued)
Batch size	64
Learning rate	1e-4 (cosine)
Warmup	2,000 steps
Augmentation	ColorJitter + RandomSharpness + RandomAffine
Cameras	3 (side_1, side_2, wrist)
Actions	7-dim (x, y, z, rx, ry, rz, gripper)
VLM layers	8 (pruned from 16)
Expert width	0.75x
Hardware	1x NVIDIA H200
Training time	~16 hours total

EdgeVLA Family

Model	Params	MSE	Cosine Sim	Gripper	Latency	HF Repo
SmolVLA	450M	0.618	0.663	94.9%	169ms	lerobot/smolvla_base
Base	363M	0.458	0.713	96.5%	162ms	enfuse/edgevla-base-fmb
Small	228M	0.515	0.679	95.8%	90ms	this repo
Tiny	164M	0.555	0.654	95.1%	57ms	enfuse/edgevla-tiny-fmb

Quick Start

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("enfuse/edgevla-small-fmb")
policy.eval()

Fine-Tuning on Your Own Data

git clone https://github.com/enfuse/edgevla
cd edgevla

python edgevla/train.py \
  --base_policy enfuse/edgevla-small-fmb \
  --dataset your_lerobot_dataset \
  --fastvit_variant fastvit_t8 \
  --num_vlm_layers 8 \
  --expert_width_multiplier 0.75 \
  --lr 3e-5 \
  --steps 50000 \
  --batch_size 64

See the training README for full configuration options and multi-round training strategy.

Attribution

Architecture from DynamicVLA (Xie et al., 2026). VLM layer pruning is our contribution. Built on SmolVLA, FastViT, and LeRobot.

@article{xie2026dynamicvla,
  title={DynamicVLA: Efficient Vision-Language-Action Model via Dynamic Fusion for Robotic Manipulation},
  author={Xie, Yue and others},
  journal={arXiv preprint arXiv:2601.22153},
  year={2026}
}