VLA-DSS
A 28.9M-parameter, RGB-only Vision-Language-Action model for robot manipulation, built for the efficiency frontier: strong performance at a fraction of the parameters of generalist VLAs.
- PyTorch
- DINOv3
- Wavelet scattering
- Fourier Neural Operator
- Robot learning
VLA-DSS is a Vision-Language-Action model for robot manipulation built around one question: how much capability can you keep while shrinking a generalist VLA by an order of magnitude? At 28.9M parameters, RGB-only, it runs in real time on weak hardware.
Architecture
- Wavelet-scattering observation encoder, a Lipschitz-stable front end that gives the policy provable robustness to small input perturbations.
- Frozen DINOv3 vision, strong visual features without the training cost.
- FiLM fusion, language conditions vision through feature-wise modulation.
- Fourier-Neural-Operator (FNO) action head, band-limited, resolution-invariant action chunks, rather than a standard MLP or diffusion head.
Results
- LIBERO, Object 79.5% (DAgger, N=200); the aux-x-y variant reaches Object 71% / Spatial 73% / Goal 72% (N=100).
- Robustness, corruption-augmentation training makes the policy blur-invariant (blur-2: 12% → 78%) at only ~5pp clean cost; noise-neutral.
- Efficiency, competitive with Octo-Small (27M) at matched size, and far smaller end-to-end than SmolVLA (450M) or Octo+T5.
The thesis is the efficiency frontier: novel, mathematically-grounded components like scattering transforms and neural operators buy robustness and resolution-invariance that let a tiny model punch well above its parameter count.