Memo, Backed by Suochen Technology, Releases Physical-WAM and RoboTwin-Phys for Embodied AI

According to a QbitAI report, embodied intelligence commercialization has run into a wall. A BeTTER benchmark jointly released by Peking University and BeingBeyond for ECCV 2026 found that the most advanced VLA models perform acceptably in standard static tests, but their success rate drops sharply once interventions such as spatial layout shifts or temporal extrapolation are introduced. A Stanford University study identified two independent failure modes in contact-rich tasks: precision failure and force failure. In other words, a VLA-driven robot only knows that it should grasp a cup when it sees one, but does not know how thin the cup wall is, whether friction is sufficient, or whether the center of gravity will shift after water is poured in. It is essentially a statistical imitator, not a participant in the physical world.

Memo focuses on the development and application of world models. It was incubated by the AVS Advanced Visual System Research Center led by Academician Gao Wen of Peking University's Information Technology Advanced Research Institute. Its core team has experience in embodied intelligence, video codec, spatial computing, and AI. Suochen Technology and Memo work together on data, computing power, and physical AI technology, and Physical-WAM and RoboTwin-Phys are the joint result of that collaboration.

Physical-WAM addresses the shortage of physical interaction data. Effective robot interaction samples come from real or high-fidelity simulated contact processes such as grasping, pushing, and pressing. Each effective interaction involves real machine motion, contact feedback, and object state changes. Such data is slow to collect, costly in hardware wear, and difficult to transfer between robot bodies. The report says usable real physical interaction samples remain in the millions, compared with billions of samples in text and image fields. Physical-WAM inserts a layer of physical token representation between perception input and action output. Traditional world models use pixels or latent vectors and can infer how the physical world evolves, but they cannot perceive physics itself: they know what the next frame will look like, but not why it changes. Physical-WAM, described as the industry's first WAM base model with physical perception, is trained on observable data and deep physical data, can detect physical changes in real time, predict the consequences of actions, and correct behavior during tasks.

The model consists of three modules: PhysLens, PhysDream, and PhysAct, forming a feedback chain from perception to prediction to generation to correction. PhysLens acts as a physical translator, extracting friction, weight, center of gravity, contact state, and other physical information that cannot be directly observed from multimodal perception signals, and outputting a unified physical token representation. PhysDream acts as a physical predictor, combining current physical representation, environment state, and candidate actions to infer future physical states, predict risks such as slipping or dropping, and provide uncertainty estimates. PhysAct handles physically conditioned action generation, bringing physical properties and future state predictions into action generation and automatically adjusting execution strategy when physical drift occurs. In one example, when a robot grasps a paper cup, PhysLens identifies the low stiffness of the cup material, PhysDream predicts that added load after pouring water may cause deformation or slipping, and PhysAct adjusts grip force and contact position, continuously correcting the strategy until the task is completed.

RoboTwin-Phys addresses a structural blind spot in existing embodied intelligence benchmarks: they measure task completion rather than physical adaptability. A model can score high on a standard test set with fixed friction coefficients and fixed object masses, yet have no robustness to perturbations in physical parameters. Under the same pushing force, a high-friction jar may tip over, while a low-friction jar may slide away. When a rolling pin is clamped at the same position, a shift in the center of gravity can cause the rod to swing. These physical drifts are common in the real world but are difficult to validate in traditional evaluations. RoboTwin-Phys is described as the first robot manipulation benchmark centered on physical factor perturbations, turning abstract physical parameters into controllable, labelable, and reproducible variables. It follows a three-layer progressive logic: whether physical attribute inference is accurate, whether tasks can still be completed under physical perturbations, and where the upper bound of performance gaps lies. Based on RoboTwin, it adds 13 categories of common real-world physical perturbations across five dimensions: object properties, contact properties, damping, task environment, and camera configuration. It supports both single-factor perturbation and multi-factor combined perturbation evaluation. Its engineering value includes providing physical ground truth such as physical values, contact force logs, object kinematics data, and failure timelines, so evaluation can ask not only whether a task succeeded but also whether the model estimated correctly and when failure began. It also uses fixed-seed paired evaluation so different models face the same perturbation instances, reducing luck, and it separates wrong physical state recognition from correct recognition with an action policy that fails to keep up. The benchmark is open sourced upon release, with project code, datasets, and leaderboard all open for third-party verification and reproduction.

The definition of a world model is still not unified. Computer vision researchers understand it as video generation, the robotics field treats it as a state predictor, and the reinforcement learning community views it as an environment proxy from the perspective of model-based reinforcement learning. A July 2026 KPMG report, "Crossing the Singularity: How World Models Reshape the Embodied Intelligence Industry and New Business Paradigms," offered a clearer coordinate system: representational world models answer what the world is, generative world models answer what the world will become, interactive world models answer what happens when the world is changed, and integrated understanding-generation-prediction world models try to answer how to understand, predict, and autonomously change the world.

Memo's Physical-WAM explores a possible technical path for physical AI, and the company says it is still expanding capability boundaries. Three open questions remain. First, where is the granularity boundary of physical tokens? Friction, mass, and center of gravity are highly abstract parameters, but real contact also involves anisotropy of surface roughness, nonlinear elastic behavior of materials, and dynamic changes in contact area. Second, how efficient is cross-embodiment physical transfer? The same object may require different action strategies for a two-finger gripper and a dexterous hand, and the mapping from physical parameters to actions depends heavily on the morphology and dynamics of the end effector. Third, where is the fidelity ceiling of simulated physics? PhysDream's inference accuracy is limited by the physical fidelity of the simulation engine and the noise level of real sensors. When the real world contains effects that simulation cannot model, such as large nonlinear deformation of fabrics or the flow of granular materials, the robustness boundary of the closed loop remains a question.

Memo's next route will focus on "vision first, force, touch, and sound progressively" to iterate multimodal physical representation models, expand physical interaction datasets, optimize model inference latency and multi-body adaptation, and work with upstream and downstream industry players to complete the three pieces of physical data, world models, and evaluation standards. "Only when robots truly understand the physical world can embodied intelligence scale from laboratories to real industrial scenarios," Memo CEO Li Minghao said. He said Memo will take an open-source and open approach to help the industry accelerate toward the "GPT moment" of embodied intelligence. The QbitAI report said the article was provided by Memo and that QbitAI was authorized to repost it, with views belonging to the original author.