OmniBrain-0 folded manipulation and navigation into one 9B decision model, and beat specialist policies on LIBERO-Plus while running 8.6x faster than GPT-5.6-Sol
AspireAI community · arXiv
OmniBrain-0 is a unified embodied agent from the Xiaomi Embodied Intelligence Team with Tsinghua that coordinates manipulation and navigation through a single high-level decision model instead of stitching separate specialists behind a tool interface. Training combines embodied pretraining, domain-specialized agent distillation, and multi-teacher on-policy distillation, plus a Rubric-Guided Recursive Self-Improvement loop where a Meta-Agent turns execution evidence into refined rubrics and verified trajectories.
The numbers are concrete. On LIBERO-Pro the 9B model reaches 65.5 percent overall success, against 78.3 percent for few-shot GPT-5.6-Sol, but average task completion time drops from 13.8 minutes to 1.6 minutes, an 8.6x speedup. On LIBERO-Plus it hits 83.8 percent, ahead of every compared skill policy and few-shot GLM-5.3-Flash at 69.3 percent. On RoboTwin 2.0 it reaches 88.2 percent overall with a 2.8x speedup over GPT-5.6-Sol and a 3.0-point success gap; RSI narrows that gap to 1.8 points. On RoboCasa365 it scores 39.6 percent, more than tripling a few-shot Qwen3.5-9B at 12.2 percent. The RSI variant raises LIBERO-Pro to 69.3 percent at the same completion time.
The honest limits: GPT-5.6-Sol still wins overall on RoboCasa365 at 86.2 percent versus OmniBrain-0's 43.2 percent with RSI, and on real-world manipulation GPT-5.6-Sol leads at an 86.7 percent macro-average against OmniBrain-0's 73.3 percent. These are the authors' own benchmark runs, and the RSI data flywheel generates its own training signal, so the improvement loop is measured on the distribution it produces.
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.