How does a brain with physical intelligence change from VLA to WBI?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

Over the past few years, the popular VLA route in the field of physical intelligence has greatly boosted the cognitive and generalized operational capabilities of robots in open environments. However, VLAs, which are trained largely on desktop or arm static baseline data, are unable to complete dozens of degrees of freedom when they enter bipedal robots with physical intelligence, are heavily driven and need to maintain momentary dynamic balance. As a goal of greater coordination, recent industry attention has begun to focus on the goal of full-body intelligent WBI to design a brain capable of truly mastering the physical.

Directory

01. How to understand the new goal of WBI with physical intelligence?
What is the relationship between WBI and WBC, VLA, and what is the industry's consensus and disagreement about WBI?...

How to understand the new goal of WBI with physical intelligence?

In the process of extending AI to the physical world, the research focus of AI is shifting from "higher semantic understanding" to "lower physical control", with its "brains" undergoing a continuous restructuring in terms of representation, control of interfaces, physical reasoning and data perception sources.

In the past few years, VLA's work as a representative of visual-language-action models has made breakthroughs in fixed-base or boundary scenarios. However, as the body-intelligent body expands into high-frequency structures such as two-legged robots, the industry has begun to pursue the "Whole-Body Intelligence" paradigm that deeply blends navigation, gestural balance and agile operation at the strategic level.[1-1] [1-2] [1-3]

1 The brain of the traditional VLA model attempts to abstract the physical world into a simple mapping of the "visual input $$\rightarrow$$$ endpoints (or trajectories) ", but it is easy to ignore the movement of the base, torsion, and focus shift.

2 In conventional architecture, disc movements or bipedal walks are separated as separate modules. Upper planners typically only give end trajectory or discrete footprint instructions, and full-body control (WBC) can only passively compensate for torque at the bottom, causing robots to fail to achieve true "full-body coordination".

3 The recent rise of full-body intelligence (WBI) emphasizes the ability of robots to understand the mechanical laws of the physical world natively and, through single or tightly coupled neural network strategies, to directly translate advanced multimodal semantic instructions into the ability to produce coherent motion that includes full-body focus adjustment, quadratic coordination, multi-touch balance, and precise force control.

3.With the goal of WBI, the industry is trying to push the shift from "local posture predictor" to "all-body physical coordination brain" to explore more coordinated "physical intelligence" systems.

1 In a speech to the RSS 2026 in mid-July, assistant professor of the University of Hong Kong and founder of Future Resources, Li Tianyan outlined the evolutionary paradigm shift from "whole-body control" to "whole-body intelligence", pointing out the need for robots in the future to achieve full-body intelligence that covers mobility, balance, contact and operation.[1-4]

2 Google DeepMind's Gemini Robotics 2 introductory document, released at the end of July, highlights the "Intelligence Layer", which consists of a combination of three models, and the full-body intelligence it brings to the robot.[1-1]

3 Current Robotics and Symbiosis Robotics have also demonstrated a breakthrough in the overall coordination of robotic systems in their latest work, attracting considerable attention.[1-5] [1-6]

How has recent work achieved WBI under a lack of consensus on technical routes?

Although "going full-body intelligence" has become a common pursuit for the next phase of the physically intelligent academic community and industry, due to existing disagreements between different teams on the selection of system architectures, the definition of data collection paradigms, and the underlying design of motion-generating engines, there are similarly different routes for how to build WBI systems.

2 Google DeepMind used a precise three-model stack architecture in its Gemini Robotics 2 work in July, combining the long-range reasoning capabilities of larger models with local high-frequency response needs.

1 The work combines the VLA model Gemini Robotics 2 with the body reasoning model Gemini Robotics ER 2 (VLM) and the lightweight VLA model Gemini Robotics On-Device 2 with three models for decision making using "ER 2" as the advanced brain to coordinate VLA execution movements to complete tasks...