Noematrix

At the Yunqi Conference, Alibaba Cloud announced a full-stack upgrade of its data and AI infrastructure aimed at the agent era, covering model production, deployment, and self-optimization. The release includes full-modal data processing in MaxCompute and EMR that cuts embodied AI data processing time by 40% and can turn petabyte-scale raw data into LeRobot or RLDS training sets, plus PAI-DLC 3.0 with Xfuse auto-tuning for 2.4x faster end-to-end multimodal training. New agent-facing services include OpenLake/Agentic Lake, ADA, PAI-InferX, PAI-RLS, Flink Streaming Agent, and PAI-CrystalLLM.

Alibaba Cloud CTO Li Feifei framed the stack around Model, Context, and Harness, while Noematrix chief scientist Lü Jun discussed embodied AI data quality and iteration cycles.

Noematrix (穹彻智能) has released RoboRSI, a multi-agent self-evolution framework aimed at real-world robot tasks. It coordinates four agents—Manager, Planner, Engineer, and Reviewer—in an execute-diagnose-revise-re-execute loop, and introduces a top-down skill refinement (TSR) four-layer skill tree to reuse capabilities.

In experiments reported by the company, code solidification lifted the LIBERO benchmark pass rate from 21.5% to 29.0% and cut median token consumption by 29.4%. Zero-shot adaptation improved from 32/120 to 95/120 tasks on LIBERO and from 9/50 to 36/50 on RoboTwin. The release notes also flag that general foundation models such as GPT-6 Astra still show safety and stability problems in real robot control.

Noematrix has released Noe-0, a pretrained model for embodied intelligence based on a world action model (WAM) architecture. Its training data consists entirely of direct human operation recordings in real environments, spanning more than 50 cities and hundreds of thousands of hours.

The model uses video prediction as its core objective, with action prediction as an auxiliary task. The company says this approach shows embodied intelligence models can be trained without traditional teleoperation data and can transfer across robot embodiments. Model size and access terms were not disclosed.