Humanoid boxing was learned in about seven minutes of single-GPU training on top of a pretrained whole-body controller
AspireAI community · arXiv (HoloBoxing authors)
HoloBoxing is a recipe rather than a new controller. It starts from a pretrained whole-body controller and separates reusable motor skills from task decisions: a compact planner requests future body-point targets while the frozen controller handles joint-level execution. A learned decoder maps 32-dimensional actions to structured target windows, and measured-state imitation initialises the planner before reinforcement learning takes over.
The point of the decomposition is that task learning never has to relearn whole-body control. It keeps optimisation focused on a compact upper policy and supplies a starting behaviour before reward-driven exploration begins. Counting from an available pretrained controller, the recipe needs about seven minutes of additional training, including reused representation fitting and a single-GPU task run - and the authors are explicit that this excludes controller pretraining, data preparation, evaluation and search.
The limitation is real: the evaluation is simulated boxing experiments with component controls, not physical robot deployment, and the seven-minute number depends entirely on the pretrained controller you already have.
https://arxiv.org/abs/2610.32909
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.