VLA-Stem added a reflex layer to vision-language-action models and doubled collision-free task success in six constrained simulation settings
AspireAI community · arXiv
Vision-language-action models are trained to manipulate objects in unconstrained settings, which leaves them exposed to spatial constraints and moving obstacles, and the obvious fix of fine-tuning on obstacle-avoidance data is expensive and can degrade the base task performance. VLA-Stem instead adds a separate lightweight safety policy that never touches the underlying VLA weights: it converts the VLA's action chunks into motion references, and a reinforcement-learned Stem policy conditioned on those references, proprioception and local obstacle information adjusts execution at inference time, leaving the high-level task intent intact. Across six constrained simulation settings the authors report 52.8 percent average collision-free task success against 23.7 percent for the strongest safety baseline, while preserving base-policy performance in obstacle-free settings, and real-world experiments show consistent collision reductions under both static and dynamic obstacles. The mechanism borrows the idea of a brainstem from biology, and it inherits the limits of its training setup: the simulation numbers are the authors' own, the real-world evidence is qualitative, and the reflex layer still has to be trained per robot and environment.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.