"Autonomous Variable Robotics" Closes 1 Billion Yuan Series A+ Round, Open-Sources End-to-End Embodied Foundation Model | Yunqi Capital

Today, X Variable Robotics, a Yunqi Capital portfolio company in embodied intelligence, announced the completion of a RMB 1 billion Series A+ funding round. The round was led by Alibaba Cloud and CAS Investment Management, with participation from China Development Bank Capital, HSG, and Zhence Capital. Existing investors including Meituan Strategic Investment, Legend Star, and Legend Capital continued their support. Yunqi Capital has backed X Variable Robotics since its Pre-A round.

At the same time, X Variable Robotics also announced the open-source release of its end-to-end embodied intelligence foundation model WALL-OSS. In this edition of Yunqi Capital, we bring you the latest on X Variable Robotics' funding and business developments.

The following content is adapted from "X Variable Robotics"

The proceeds from this round will be used for continued training of X Variable's fully self-developed general-purpose embodied intelligence foundation model and R&D iteration of hardware products.

Since its founding in late 2023, X Variable has established an end-to-end unified large model as its technical path toward general-purpose embodied intelligence, and recently released its fully self-developed wheeled dual-arm humanoid robot controlled by a multimodal large model — Quanta X2. X Variable's integrated software-hardware development approach, along with its forward-looking technical vision and achievements, has gained recognition from national investment platforms, top-tier domestic and international investment institutions, and industry capital.

As the earliest company in China to achieve an end-to-end embodied intelligence large model, X Variable independently developed the "WALL-A" series of VLA (Vision-Language-Action) manipulation large models, constructing a unified cognition and action framework. Within a unified representation space, the model simultaneously processes perception, reasoning, and action, directly performing cross-modal causal reasoning and action decision-making — enabling robots to ultimately think and work like humans. Currently, the "WALL-A" model has demonstrated zero-shot generalization capabilities on certain entirely novel task types it was never trained on.

At the same time, the company pioneered an end-to-end embodied chain-of-thought reasoning framework, performing deep reasoning based on multimodal inputs and generating multimodal outputs, forming a complete closed loop of autonomous decision-making, execution, exploration, and reflection. The model tightly integrates language understanding, visual perception, and action execution, forming a reasoning process closer to human thinking. It has successfully broken through the bottleneck of multi-step, long-sequence tasks, substantially improving task completion rates and greatly expanding the boundaries of robots' ability to handle complex real-world scenarios.

By mid-this year, the company first achieved complex manipulation using an embodied intelligence large model to control a high-DOF dexterous hand. Previously, X Variable released a video of its self-developed large model controlling a high-DOF dexterous hand to delicately pick up and distribute playing cards — elastic, easily deformable objects.

Currently, to advance research and application of embodied intelligence large models, X Variable has open-sourced its developer-facing embodied foundation model "Wall-OSS" and publicly released relevant training code, enabling developers worldwide to rapidly fine-tune and deploy it on their own robot platforms.

Wall-OSS features powerful generalization and reasoning capabilities, outperforming other foundation models on long-horizon manipulation tasks. As a multimodal base model, it also demonstrates strong causal reasoning, spatial understanding, and reflection capabilities. Wall-OSS is an open-source embodied foundation model trained on large-scale real-world data.

In terms of model architecture, it innovatively designs a "shared attention + expert routing (FFN)" architecture, enabling lossless knowledge transfer from VLM to manipulation models and achieving deep coupling of language and action. In training methodology, it pioneers a three-stage training paradigm of "first discrete, then continuous, then joint," ensuring stable, lossless transfer and extension of VLM's cognitive capabilities to physical actions. Additionally, unified cross-level chain-of-thought enables forward arbitrary mapping across levels of abstraction, allowing the model to seamlessly switch between high-level decision-making and low-level execution within a single differentiable framework.

On the hardware side, in August this year, X Variable released its fully self-developed wheeled dual-arm humanoid robot Quanta X2. In less than six months, the company achieved full-stack self-development of robot bodies, high-DOF dexterous hands, and exoskeleton teleoperation data collection devices.

Quanta X2 is a model-native general-purpose robot platform. Its design not only addresses model training and complex manipulation task requirements, but also achieves comprehensive balance and optimization across payload capacity, operational workspace, movement speed, and control precision — core metrics.

Quanta X2's five-finger dexterous hand employs a biomimetic structural design, with 20 degrees of freedom per hand and the ability to sense subtle pressure changes. Meanwhile, based on integrated arm-hand exoskeleton technology, X Variable pioneered an industry-leading "humanoid robotic arm + high-DOF dexterous hand" integrated whole-body teleoperation solution. Quanta X2 not only collects high-quality data to feed back into model training, but will also be deeply integrated with self-developed models to truly enter and deploy in real-world scenarios.

With the improvement of integrated software-hardware capabilities, X Variable's robots have already established partnerships with leading service and industrial clients and been deployed across multiple scenarios. Going forward, X Variable will work with clients to co-build an open ecosystem around models and hardware, driving further advancement of embodied intelligence.