July 22, 2026

Xiaomi Robotics presents a vision-language-action foundation model pretrained on over 100,000 hours of real-world manipulation trajectories using an automatic natural-language scene-change annotation pipeline, then aligned to specific robot platforms. It achieves 57.6% success on RoboCasa365 (vs. prior best 46.6%) and a score of 20.07 on RoboDojo (vs. prior best 13.07), with code and weights promised for release.

DAMO Academy's RynnBrain 1.1 is a family of embodied foundation models (2B, 9B, 122B-A10B parameters) unifying spatio-temporal and physics-grounded reasoning for perception, spatial reasoning, localization, and planning, adding contact-point prediction and native 3D grounding for smaller variants. The companion RynnBrain-VLA introduces a cross-embodiment action space deployed across three distinct robot platforms, with the 122B model outperforming existing proprietary and open-source baselines.

TimeLens2 predicts variable-cardinality sets of evidence intervals identifying when events occur in video, built on a new TimeLens2-93K supervision dataset and a novel Temporal Wasserstein reward that gives matching-free feedback for unequal prediction cardinalities. The 2B/4B/8B model variants improve 14.2-18.1 mIoU points over their Qwen3-VL backbones across seven benchmarks.

RAGU is an open-source modular GraphRAG system that separates entity extraction from consolidation via two-stage typed extraction, DBSCAN-backed deduplication, LLM summarization, and Leiden community detection. Its accompanying 7B model, Meno-Lite-0.1, outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% relative harmonic mean) despite being far smaller, and the system achieves higher evidence recall (up to 0.84 vs. <=0.76) than competitors on medical-domain benchmarks; it runs on a single GPU and is MIT-licensed with public code and weights.