The embodied AI business is quickly altering its priorities, with software program intelligence rising as the most important aggressive benefit over {hardware} enhancements.
According to a report by Pandaily, builders unveiled 13 new embodied AI basis fashions and world fashions throughout June 2026, averaging one launch each 48 hours. The tempo displays a rising business give attention to enabling robots to assume, purpose, and adapt extra successfully quite than merely enhancing their mechanical capabilities.
The shift comes as robotics corporations more and more view AI fashions as the important thing expertise figuring out how effectively humanoid robots and autonomous machines carry out in real-world environments.
BAAI focuses on world understanding
Rather than following a single method, main AI labs and robotics corporations are focusing on particular weaknesses that proceed to restrict robots. At the 2026 Zhiyuan Conference, the Beijing Academy of Artificial Intelligence (BAAI) launched two world-model tasks.
The first, Wujie Physis-v0.1, is designed to foretell the subsequent bodily state of an setting by combining a number of kinds of sensory data, together with video, RGB-D knowledge, 3D level clouds, and force-tactile alerts, right into a shared latent illustration.
BAAI additionally unveiled Wujie RoboBrain Orca, which capabilities as a robotic mind by combining language and visible data with causal reasoning and multimodal decoding.
Alibaba bets on higher mannequin design as an alternative of larger scale
Alibaba took a unique path with the launch of its Qwen-Robot household on June 16.
Rather than relying totally on bigger datasets and extra robotic coaching, the corporate argued that the range of real-world environments can’t be solved by means of scale alone. Instead, it proposed model-level alignment strategies tailor-made for various robotic duties.
The suite contains:
- Qwen-RobotNav, which adjusts visible consideration for robotic navigation.
- Qwen-RobotManip, which standardizes state and motion areas for manipulation duties.
- Qwen-RobotWorld, which predicts world dynamics utilizing natural-language motion interfaces.
New fashions goal particular robotic weaknesses
Several different organizations launched fashions geared toward fixing particular person bottlenecks that proceed to restrict robotic efficiency.
CasiaHand unveiled Brain-Si 0.5, described because the world’s first human-like dexterous manipulation mannequin. It makes use of a three-layer structure that mixes planning, manipulation capabilities, and bodily interpretable fashions for duties equivalent to greedy, handovers, bimanual coordination, and human-robot interplay.
GalaxyBot launched AstraBrain-WBC 0.5, a cerebellum-inspired basis mannequin for whole-body humanoid management. Built on a GPT-style causal Transformer structure, it was skilled on roughly 2 billion frames of human motion knowledge and incorporates 80 million parameters.
Other releases included RoboScience’s Visics structure, which separates world fashions from operation fashions utilizing object trajectory representations; Current Robotics’ Curl-0, which approaches whole-body dexterous manipulation by means of coupled coaching; and BoundlessPower’s MWA world mannequin for modeling long-sequence bidirectional bodily causality.
The focus is shifting from functionality to reliability
Pandaily famous that regardless of their completely different technical approaches, these tasks share a typical aim. Instead of merely showcasing spectacular robotic demonstrations, researchers are more and more attempting to grasp why robots nonetheless wrestle with many on a regular basis duties.
Teams are focusing on completely different bottlenecks, together with tactile sensing, whole-body coordination, transferring expertise from simulation into the true world, and planning over lengthy sequences of actions. The rising emphasis is on constructing AI architectures that may overcome these limitations quite than…
