Open Source
SimVLA: Attributing Gains in VLA Models Through Controlled Ablation
arXiv cs.RO0 views
arXiv:2602.18224v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models often bundle architectural changes with different pretraining data, backbone scales, and optimization recipes, obscuring what actually drives progress. We introduce SimVLA, a deliberately minimal VLA, a standard vision-language backbone with a lightweight continuous-action head, as a controlled testbed for this attribution problem; our contribution is methodological and empirical (a controlled attribution study and a minimal-baseline calibration), not a new technique. Across LIBERO, CALVIN, WidowX, and Google Robot, one-knob-at-a-time ablations show that training dynamics are the largest measured driver, producing performance swings of $\Delta$$\approx$54--94\%, while task configuration and architecture have smaller effects under matched settings. Despite using a small backbone and no additional large-scale robot-data VLA pretraining before benchmark-specific training, SimVLA reaches 98.6\% average success on LIBERO, remains competitive across the other benchmarks, and transfers to held-out real-robot scenes. These results suggest that simple, carefully controlled VLA baselines remain underexplored and provide a calibration point for future architectural claims within the VLM-encoder plus lightweight-head family. Website: https://frontierrobo.github.io/SimVLA
Source:https://arxiv.org/abs/2602.18224PreviousLCLA: Language-Conditioned Latent Alignment for Vision-Language NavigationNextAn Autonomous Sea Turtle Robot that Follows Marine Animals
Related posts
A sticky Kakeya set over F3 that beats the classical construction, and where the question stands
0dev.to
In 2022 I learned to build Android apps in a country where I could not publish one
0dev.to
Are You Buying a Company or a model_id?
0dev.to Discuss
How to Test a Resistor, Diode, LED and Capacitor with a Multimeter (Before You Blame Your Code)
0dev.to Beginners
Omni-modal decomposition autoencoders learn full-stack wearable disentangled representations
0arXiv eess.SP
Spectral Vision Transformer for Efficient Tokenization with Limited Data
0arXiv eess.SP